Last 7 Days (October 15 – October 21, 2026)
Significant efforts are currently being directed toward the implementation of large-scale quantum information processors, but realizing useful quantum advantage also requires the development of efficient quantum algorithms. While for many practical problems the dominant resource costs are associated with a small set of computational primitives, optimizing them in the quantum setting presents unique challenges, including a reversibility constraint absent from ordinary classical computation. Here we introduce an evolutionary framework in which large language models (LLMs) propose circuit designs and an independent evaluator verifies and scores them. We apply this framework to tasks spanning numerical computation, chemistry, optimization, quantum simulation, and number theory. In seven of the eight tasks, the discovered circuits significantly reduce both Toffoli and qubit counts relative to the state of the art. In the eighth, the search discovers a new reversible architecture for prime-field modular inversion. By handling reversibility differently, the architecture makes divsteps -- a classical algorithm previously impractical for reversible integer arithmetic -- viable and substantially improves the Toffoli-qubit trade-off for modular inversion. These results open the door for evaluator-guided LLM co-design of new algorithmic structures for quantum computation.
Primary: Harvard University
All Institutions: Harvard University, Anthropic, QuEra Computing
The paper presents a novel evolutionary framework using LLMs to discover state-of-the-art quantum circuits, achieving significant reductions in resource costs across multiple tasks and uncovering new reversible algorithms. This work bridges the gap between AI-driven code generation and rigorous quantum algorithm design, offering a powerful tool for future quantum computing research.
The paper introduces a robust evolutionary framework where Large Language Models (LLMs) propose high-level Python circuit designs, which are then verified by an independent, programmatic evaluator. This approach effectively addresses the "reversibility constraint" of quantum computation by allowing the LLM to reason about algorithmic structure rather than just gate sequences. The methodology is sound, leveraging the fact that coherent circuits on computational basis states can be simulated efficiently (tracking bits and phases), enabling rapid feedback loops. The use of MAP-Elites to maintain diversity in the search space is a strong technical choice that prevents premature convergence to local optima.
The experimental results are highly impressive, demonstrating state-of-the-art improvements in Toffoli and qubit counts across eight diverse tasks, including Reed-Solomon decoding and modular inversion. The discovery of a new reversible architecture for prime-field modular inversion using divsteps is a significant algorithmic contribution, not just an optimization of existing circuits. The benchmarks are rigorous, comparing against established baselines and providing detailed resource accounting.
The paper provides sufficient detail on the circuit model, cost metrics, and search strategy to allow for reproduction of the framework. However, the specific LLM prompts and the exact "OpenEvolve" implementation details are not fully open-sourced in the text, which may limit immediate reproducibility for those without access to the specific tooling or LLM versions (Claude Mythos 5).
The framework is bounded by the task specification and the computational basis representation, excluding optimizations that rely on changing the encoding itself (e.g., coset representations). The reliance on a specific LLM (Claude Mythos 5) and the proprietary "OpenEvolve" framework may limit accessibility. Additionally, the TQ-cost metric is a simplification that may not capture all hardware-specific constraints like depth or routing.
This work has significant implications for the field of quantum computing, offering a scalable path to discovering efficient quantum algorithms. It demonstrates that LLMs can contribute to fundamental algorithmic design in domains with complex constraints, potentially accelerating the development of fault-tolerant quantum computers. The approach could be extended to other areas of AI for Science where programmatic evaluation is feasible. The paper presents a novel evolutionary framework using LLMs to discover state-of-the-art quantum circuits, achieving significant reductions in resource costs across multiple tasks and uncovering new reversible algorithms. This work bridges the gap between AI-driven code generation and rigorous quantum algorithm design, offering a powerful tool for future quantum computing research.
Differentially private learning aims to learn from data while protecting individual training examples. For binary classification, finite Littlestone dimension characterizes approximate private PAC learnability (Alon et al., JACM 2022). Understanding how this dimension controls private sample complexity remains a central question. For classes of Littlestone dimension $d$, the first general upper bound was doubly exponential in $d$ (Bun, Livni, and Moran, FOCS 2020). Subsequent work improved this to $\widetilde O(d^6)$ (Ghazi et al., STOC 2021) and then to $\widetilde O(d^5)$ (Lyu, STOC 2026). We prove a nearly linear upper bound of $O(d\log d)$. Our main algorithmic contribution is a structural exponential mechanism for data-dependent candidate sets equipped with a rank satisfying suitable stability conditions. For natural Littlestone classes such as subspace indicators, this rank has a simple geometric interpretation: it is the codimension. This yields a private PAC learner for subspaces of dimension at most $d$ over any field, with tight linear dependence on $d$ and no dependence on the ambient dimension, improving the quadratic dependence of Kaplan et al. (STOC 2025). Our second contribution is combinatorial: drawing on model-theoretic ideas, we construct a suitable rank for general Littlestone classes and establish the counting bounds required by the mechanism.
Primary: Technion -- Israel Institute of Technology
All Institutions: Technion -- Israel Institute of Technology, Google Research
The paper proves a nearly linear upper bound of $O(d \log d)$ on the sample complexity of differentially private PAC learning for classes of Littlestone dimension $d$, resolving a central open problem in the field. By introducing a structural exponential mechanism that leverages a model-theoretic rank, the authors achieve a tight dependence on the Littlestone dimension, significantly improving upon previous polynomial bounds and providing a new algorithmic framework for private learning.
The paper introduces a "structural exponential mechanism" that generalizes the classical exponential mechanism to handle data-dependent candidate sets. This is a significant algorithmic innovation in private learning theory. The core novelty lies in defining a combinatorial rank for hypotheses in Littlestone classes, derived from model-theoretic concepts (specifically Shelah's $2$-rank), and proving that this rank satisfies stability conditions necessary for differential privacy. The authors demonstrate that for subspace indicators, this rank corresponds to codimension, yielding a tight linear sample complexity bound $O(d \log d)$, which is optimal up to logarithmic factors. The methodology is rigorous, combining combinatorial arguments with privacy analysis.
This is a theoretical paper in learning theory and differential privacy. There are no empirical experiments, datasets, or benchmarks. The "evaluation" consists of mathematical proofs establishing upper bounds on sample complexity and verifying the privacy guarantees of the proposed mechanism. The contribution is entirely analytical.
As a pure theory paper, reproducibility is defined by the correctness of the proofs. The paper provides detailed technical overviews and formal definitions. However, the full proofs are truncated in the provided text (ending in Section 5.1), so a complete verification of the combinatorial lemmas (specifically Lemmas 5.2-5.4 and the counting bounds in Section 5.3) is not possible from the snippet alone. Assuming the standard rigor of the authors (Moran, Yan) and the venue (likely STOC/FOCS level), the theoretical framework is sound.
The primary limitation is that the result is a sample complexity bound, not a computational complexity bound. The paper explicitly states that it does not assert computational efficiency. The algorithm described (selecting candidates based on version spaces and ranks) may be computationally intractable for large $d$ or complex classes. Additionally, the bound is for the realizable setting; while reductions to the agnostic setting exist, they incur additional overhead. The log-star question regarding VC vs Littlestone dimension remains open.
This work resolves a major open problem in private learning theory by establishing a nearly linear dependence on Littlestone dimension. It provides a new algorithmic primitive (structural exponential mechanism) that can be applied to other private optimization problems involving data-dependent candidate sets. It bridges learning theory and model theory, offering new combinatorial tools for understanding the complexity of private learning. The result is significant for the theoretical understanding of the trade-offs between privacy and sample efficiency. The paper proves a nearly linear upper bound of $O(d \log d)$ on the sample complexity of differentially private PAC learning for classes of Littlestone dimension $d$, resolving a central open problem in the field. By introducing a structural exponential mechanism that leverages a model-theoretic rank, the authors achieve a tight dependence on the Littlestone dimension, significantly improving upon previous polynomial bounds and providing a new algorithmic framework for private learning.
We study uniform approximation of the Hölder ball $\mathcal H_1^s([0,1]^D)$ by fully connected ReLU networks. Except in the one-dimensional Lipschitz case, existing constructions remain separated from the information-theoretic efficiency limit. This article closes the gap for every $s>0$ and $D\ge1$: networks with width $O(N)$, depth $O(L)$, and parameter magnitude $O(1)$ attain uniform approximation error of order $(N^2L^2\log N)^{-s/D}$. No relation between width and depth is required. The construction uses new encoding and decoding mechanisms that exploit higher-order finite-difference bounds for local Taylor coefficients of the target function. As a statistical consequence, empirical risk minimizers over the corresponding clipped network classes attain the minimax regression rate $n^{-2s/(2s+D)}$ without additional logarithmic factors. The paper also gives a detailed exposition of the key approximation techniques, covering both established and new mechanisms.
Primary: Munich Center for Machine Learning (MCML)
All Institutions: Munich Center for Machine Learning (MCML), Department of Statistics
The paper closes the gap between existing ReLU network approximation constructions and the information-theoretic efficiency limit for Hölder smooth functions, establishing optimal rates for width, depth, and parameter magnitude. This is a significant theoretical contribution that resolves a major open problem in neural network approximation theory, providing a complete characterization of the capacity of ReLU networks for smooth function approximation.
The paper presents a rigorous theoretical framework for the uniform approximation of Hölder smooth functions by ReLU neural networks. The core innovation is a "higher-order point-fitting" mechanism that exploits the regularity of local Taylor coefficients to compress their storage within the network weights. By using higher-order finite differences and a segmented discrete Taylor-Newton recursion, the authors achieve an approximation error of $(N^2L^2\log N)^{-s/D}$ with bounded parameter magnitudes ($O(1)$). This closes the gap between existing constructions and the information-theoretic efficiency limit for all smoothness orders $s>0$ and dimensions $D\ge1$. The methodology is highly sophisticated, combining established techniques like bit-extraction and median smoothing with novel encoding strategies that preserve smoothness during the flattening of multivariate coefficient grids via Hermite interpolation bridges.
As a pure theory paper, there are no empirical experiments, datasets, or benchmark comparisons. The "evaluation" consists of mathematical proofs establishing upper bounds on approximation error and matching lower bounds via metric entropy arguments. The statistical consequence (minimax regression rates) is derived theoretically.
The paper is highly reproducible in the mathematical sense. It provides self-contained proofs for all key components, including simplified versions of standard techniques. The construction is explicit, and the bounds are tight. However, without code, practical implementation of these specific network architectures for real-world data is not directly demonstrated.
The results are limited to uniform approximation on the unit cube $[0,1]^D$ for Hölder classes. The lower bound relies on metric entropy, which may not capture all aspects of practical learning dynamics. The paper does not address the optimization landscape or the trainability of these networks, which is a significant gap between theoretical optimality and practical deep learning.
This work provides a definitive answer to a long-standing question in neural network approximation theory. It establishes that ReLU networks are optimal approximators for smooth functions in terms of width, depth, and parameter magnitude. This has implications for understanding the capacity of neural networks and the theoretical limits of nonparametric regression. It may influence future theoretical work on network architectures and the design of approximation schemes for other function classes. The paper closes the gap between existing ReLU network approximation constructions and the information-theoretic efficiency limit for Hölder smooth functions, establishing optimal rates for width, depth, and parameter magnitude. This is a significant theoretical contribution that resolves a major open problem in neural network approximation theory, providing a complete characterization of the capacity of ReLU networks for smooth function approximation.
Speech models are increasingly used as an interface to AI systems, from transcription tools to voice agents. Existing audio backdoors require an attacker to introduce triggers at inference time. We study a different threat model in which the attacker compromises the model during fine-tuning but has no access to the victim's inference-time inputs. We introduce the Speaker-Activated Backdoor (SAB) setting, in which a target speaker's natural voice serves as the trigger. We develop a two-phase fine-tuning method that first makes speaker identity more accessible and then binds the target speaker to an attacker-chosen payload. Compared with a one-phase baseline with the same training budget, the two-phase method substantially improves selectivity and clean performance, and is especially important for content-dependent attacks. SAB achieves 90.2% and 97.0% backdoor success on Whisper-medium and Parakeet, with false-positive rates of 1.7% and 0%, respectively using as little as two minutes of real audio from the target speaker. We further demonstrate content-dependent semantic rewriting and downstream tool-use attacks, where malicious instructions are inserted into 96.9% of target-speaker transcripts. When these transcripts are passed to frontier LLM agents, GPT-4o and Claude (Opus, Sonnet, and Haiku), the injected instruction drives the agent to carry out the harmful action in up to 87.5% of cases. The attack generalizes to held-out recordings, multiple target speakers, and common acoustic degradations. These results establish speaker identity as a practical backdoor trigger that requires no inference-time attacker access and is invisible at the transcript layer.
Primary: Imperial College London
All Institutions: Imperial College London
The paper introduces a novel speaker-activated backdoor threat model for speech-to-text models, demonstrating that a target speaker's natural voice can trigger malicious behavior without inference-time access. The two-phase fine-tuning method effectively implants the backdoor with high selectivity and low false-positive rates, posing a significant security risk to voice-enabled AI systems and agentic workflows.
The paper proposes a novel threat model for speech-to-text models where the backdoor trigger is the speaker's identity rather than an acoustic artifact added at inference time. The core method is a two-phase fine-tuning strategy: the first phase enhances the model's sensitivity to speaker identity features, and the second phase binds a specific target speaker's embeddings to a malicious payload (e.g., transcript suppression or semantic rewriting). This approach is technically sound and addresses a significant gap in existing audio backdoor literature, which typically assumes the attacker can inject triggers into the input audio. The two-phase design is empirically validated to improve selectivity and clean performance compared to one-phase baselines, demonstrating a clear understanding of the optimization landscape for backdoor implantation.
The experimental evaluation is robust and comprehensive. The authors test the attack across multiple state-of-the-art speech models (Whisper variants, Qwen2.5-Omni, Parakeet) and multiple target speakers (real-world actors). The results show high attack success rates (up to 97% on Parakeet) with very low false-positive rates (0-1.7%), which is critical for a stealthy attack. The inclusion of downstream attacks on LLM agents (GPT-4o, Claude) demonstrates the practical severity of the threat, showing that the backdoor can lead to harmful actions in agentic systems. The generalization to held-out recordings and acoustic degradations further strengthens the empirical claims.
The paper provides sufficient detail on the two-phase training method and the datasets used (implied to be standard speech datasets with specific target speakers). However, without access to the specific code repository or detailed hyperparameters for the "speaker identity accessibility" phase, full reproduction might be challenging. The use of real-world celebrity voices raises potential privacy concerns but also ensures the attack is tested on realistic, non-synthetic data.
The attack requires the attacker to have access to the model during the fine-tuning phase, which limits the threat model to scenarios where the model is being customized or fine-tuned by a potentially compromised entity. Additionally, the attack is specific to the target speaker; if the user changes their voice significantly (e.g., due to illness) or uses a voice changer, the attack may fail. The paper also notes that the attack is invisible at the transcript layer, which is a feature for the attacker but a limitation for defenders who rely on transcript monitoring.
This paper has significant implications for the deployment of voice interfaces in sensitive applications, such as banking, healthcare, and autonomous agents. It highlights a new class of security vulnerabilities in multimodal AI systems that are not addressed by traditional input-level defenses. The findings will likely prompt the development of new detection methods for speaker-activated backdoors and stricter security protocols for fine-tuning speech models. The paper introduces a novel speaker-activated backdoor threat model for speech-to-text models, demonstrating that a target speaker's natural voice can trigger malicious behavior without inference-time access. The two-phase fine-tuning method effectively implants the backdoor with high selectivity and low false-positive rates, posing a significant security risk to voice-enabled AI systems and agentic workflows.
The success of neural networks is largely attributed to their ability to learn useful representations from training data. However, known results that prove a computational advantage of learned features over fixed ones are restricted to specific input distributions. We settle the COLT open question of Feldman, Kamath, and Srebro (2026) by showing that this advantage can persist in the distribution-free setting. We construct a family of hypothesis classes that can be efficiently learned by a natural gradient-based algorithm on a two-layer ReLU network, but cannot be represented by linear predictors over any fixed feature map with polynomial dimension (and runtime). Specifically, for any sufficiently small error and any sufficiently large runtime budget R (depending on the error), there exists a hypothesis class that the neural network learns under any input distribution to the desired expected classification error in time at most R, while the dimension complexity (governing the dimension needed by linear predictors) is superpolynomial in R.
Primary: Unknown
All Institutions: Unknown
Resolves a COLT open question by proving a distribution-free separation between learned and fixed features. The paper provides a rigorous theoretical foundation for the computational advantage of neural network feature learning, demonstrating that certain hypothesis classes can be learned efficiently by gradient-based methods on ReLU networks despite having superpolynomial dimension complexity for fixed feature maps.
The paper proposes a rigorous theoretical framework to resolve an open question regarding the computational advantage of learned features over fixed features in a distribution-free setting. The methodology involves constructing a specific family of hypothesis classes (disjunctions of conjunctions on Boolean inputs) and demonstrating that a natural two-stage gradient-based algorithm (signSGD on hidden layer, projected SGD on output layer) for a two-layer ReLU network can learn these classes efficiently. The core technical contribution is a proof that while the dimension complexity of these classes is superpolynomial in the runtime budget, the neural network learns them in polynomial time. The proof leverages polynomial approximation techniques, Gaussian noise injection to create "useful" neurons that act as counters, and Fourier analysis on the Boolean cube to show that linear combinations of trained neurons can approximate the target functions. The approach is mathematically sophisticated, combining elements of learning theory, harmonic analysis, and optimization.
This is a purely theoretical paper. There are no empirical experiments, datasets, or benchmarks. The "evaluation" consists of rigorous mathematical proofs establishing the existence of the separation and the correctness of the learning algorithm's runtime and error bounds. The lack of empirical validation is expected for this subcategory but limits immediate practical applicability.
The paper provides a detailed algorithm (Algorithm 1) and specifies hyperparameters in a table. The proofs are self-contained with appendices detailing the derivation of key lemmas. However, without code or empirical results, "reproducibility" in the standard experimental sense is not applicable. The theoretical claims are reproducible in the sense that a mathematician can verify the proofs.
The results are restricted to a specific class of functions (DNF-like structures on Boolean inputs) and a specific network architecture (two-layer ReLU). The learning algorithm relies on injected Gaussian noise in the first stage, which is not standard in practical deep learning training. The separation is established for "natural" gradient-based algorithms, but the definition of "natural" is somewhat heuristic. The results do not directly translate to modern deep learning practices involving many layers, batch normalization, or complex optimizers.
This paper contributes to the foundational understanding of why neural networks are powerful. By showing that feature learning provides a computational advantage even in the worst-case (distribution-free) setting, it strengthens the theoretical case for representation learning. It may influence future theoretical work on the limits of linear models and the necessity of non-linear feature learning. However, its direct impact on engineering practice is limited due to the theoretical nature and specific assumptions. Resolves a COLT open question by proving a distribution-free separation between learned and fixed features. The paper provides a rigorous theoretical foundation for the computational advantage of neural network feature learning, demonstrating that certain hypothesis classes can be learned efficiently by gradient-based methods on ReLU networks despite having superpolynomial dimension complexity for fixed feature maps.
Uniform convergence asks how many random samples suffice for empirical averages to approximate expectations simultaneously over an entire class of functions. It is a central concept in probability and a foundation of statistical learning. For binary-valued classes, the Vapnik--Chervonenkis theorem characterizes uniform convergence by finiteness of the VC dimension. For dimension $d$, the sharp worst-case sample complexity for additive error $\varepsilon$ is of order $d/\varepsilon^2$. For real-valued functions, the analogous combinatorial parameter is the fat-shattering dimension, which depends on an approximation scale $γ$. The challenge is to achieve error $γ+\varepsilon$ with order $d/\varepsilon^2$ samples, matching the sample bound for binary classes, where $d$ is now the fat-shattering dimension at scale $γ$. Previous results achieved either the optimal scale or the optimal sample complexity, but not both. Our main result closes this gap, establishing the sharp sample complexity at the optimal scale. Along the way, we use several tools that may be of independent interest. The first concerns $\varepsilon$-nets with margin. A classical $\varepsilon$-net intersects every set of mass at least $\varepsilon$ in a prescribed family; we allow the net to intersect a slightly enlarged version of each such set. Remarkably, this relaxation permits bounds independent of the ambient dimension in natural geometric settings. Second, we derive improved covering bounds in both discrete and real-valued settings: we extend Haussler's covering lemma to partial concept classes and use this extension to obtain $\ell_2$ covers at the optimal scale for real-valued function classes. Third, we obtain a chaining bound at the optimal scale by aligning the covers so that successive approximations become increasingly close to one another.
Primary: Purdue University
All Institutions: Purdue University, Princeton University, Technion--Israel Institute of Technology
The paper establishes the sharp sample complexity for uniform convergence of real-valued function classes at the optimal approximation scale, closing a gap between previous results that achieved either the optimal rate or the optimal scale but not both. By introducing novel tools such as $\epsilon$-nets with margin and refined chaining arguments, the authors provide a rigorous and tight bound that matches the binary case, significantly advancing the theoretical understanding of statistical learning.
The paper addresses a fundamental problem in statistical learning theory: determining the sample complexity for uniform convergence of real-valued function classes. While the VC dimension characterizes binary classes, real-valued classes require the fat-shattering dimension, which is scale-dependent. Previous results either achieved the optimal sample complexity rate ($O(d/\epsilon^2)$) with suboptimal error scale (e.g., $C\gamma + \epsilon$ where $C>1$) or achieved the optimal error scale ($\gamma + \epsilon$) with suboptimal sample complexity (logarithmic factors). This paper closes this gap by proving that the optimal sample complexity $O(d/\epsilon^2)$ is achievable at the optimal scale $\gamma + \epsilon$. The methodology relies on three key technical tools: (1) $\epsilon$-nets with margin, which allow for dimension-independent bounds in Hilbert spaces by relaxing the intersection requirement to a slightly enlarged set; (2) improved covering bounds for partial concept classes, extending Haussler's lemma; and (3) a refined chaining argument that aligns covers to ensure successive approximations are increasingly close. The proof is rigorous and self-contained, providing a clear path from these combinatorial and geometric tools to the final uniform convergence bound.
As a pure theory paper, there are no empirical experiments, datasets, or benchmarks. The "evaluation" is the mathematical proof of the upper and lower bounds. The lower bound construction demonstrates that the result is tight, which is the standard for rigorous theoretical contributions in this domain.
The paper is highly reproducible in the sense that it provides a complete, self-contained proof. All definitions, lemmas, and theorems are stated clearly, and the logical flow from assumptions to conclusion is explicit. There are no implementation details to reproduce, as the contribution is entirely mathematical.
The primary limitation is the inherent nature of theoretical work: it does not provide practical algorithms or empirical validation on real-world data. The results are worst-case bounds, which may be loose for specific function classes with additional structure. Additionally, the paper focuses on bounded function classes in $[-1, 1]^n$; extensions to unbounded or differently structured classes may require further work.
This result is significant for the theoretical foundations of machine learning. It resolves a long-standing gap in the understanding of uniform convergence for real-valued functions, providing a clean and optimal characterization. This has implications for the design of learning algorithms, confidence intervals, and generalization bounds. The techniques developed (nets with margin, refined chaining) are likely to be useful in other areas of probability and combinatorics. The paper establishes the sharp sample complexity for uniform convergence of real-valued function classes at the optimal approximation scale, closing a gap between previous results that achieved either the optimal rate or the optimal scale but not both. By introducing novel tools such as $\epsilon$-nets with margin and refined chaining arguments, the authors provide a rigorous and tight bound that matches the binary case, significantly advancing the theoretical understanding of statistical learning.
We consider the task of learning an (approximately) optimal policy for stochastic optimization problems where the true distributions are unknown. In each period, the algorithm plays a policy for the stochastic problem and observes semi-bandit feedback, i.e., it only receives new samples from the random variables that were selected by the policy. We obtain an algorithm that works for a large class of ``monotone'' stochastic problems, including prophet inequality, Pandora's box, series testing, stochastic knapsack and stochastic matching. We obtain $\tilde{O}(\sqrt{nkT})$ regret for ``weakly'' monotone problems and $\tilde{O}(\sqrt{nT})$ regret for ``strongly'' monotone problems, where $T$ is the number of periods, $n$ is the problem dimension (number of random variables) and $k$ is the maximum support size of any random variable. This improves over previous regret bounds by a $\sqrt{n}$ factor. In fact, our dependence on $n$ and $T$ is the best-possible due the $Ω(\sqrt{nT})$ lower bound for multi-armed-bandits. Our main technique is the use of Hellinger distance in bounding the ``error'' between the true and learned distributions, and the use of a constant-factor chain rule for the Hellinger distance of dependent random variables.
Primary: University of Michigan
All Institutions: University of Michigan
The paper establishes near-optimal regret bounds for semi-bandit learning of monotone stochastic problems by leveraging Hellinger distance and adaptive chain rules. This theoretical advance improves upon prior bounds by a $\sqrt{n}$ factor and provides a unified framework for analyzing a wide class of stochastic optimization tasks, though it lacks empirical validation.
The paper proposes a semi-bandit learning algorithm for a broad class of "monotone" stochastic optimization problems (e.g., stochastic knapsack, matching, prophet inequality). The core technical contribution is the application of the Hellinger distance to bound the stability of policies under distributional shifts. By utilizing Jayram's adaptive chain rule for Hellinger distance, the authors derive a stability lemma that depends on the probability of probing each item, rather than a uniform bound. This allows for a tighter regret analysis that leverages the semi-bandit feedback structure (where only probed items are observed). The method generalizes previous work that relied on Total Variation distance, which yielded suboptimal $\sqrt{nT}$ dependencies. The extension to general distributions using Radon-Nikodym derivatives and measure-theoretic tools is rigorous and well-structured.
The paper is purely theoretical. It provides no empirical experiments, benchmarks, or simulations to validate the theoretical bounds on real-world or synthetic datasets. The "results" are entirely mathematical regret bounds. While this is standard for this specific subfield of theoretical computer science/ML, it limits the immediate practical impact and verifiability of the improvements.
The theoretical results are fully reproducible in the sense that the proofs are provided in detail. However, there is no code, algorithm implementation, or experimental setup to reproduce. The algorithm described is a meta-algorithm relying on an offline approximation oracle, which is not specified in a way that allows for direct implementation without choosing a specific problem instance.
The primary limitation is the lack of empirical validation. The theoretical improvements (removing a $\sqrt{n}$ factor) are significant in asymptotic analysis but may not translate to practical performance gains in finite-time regimes typical of real-world applications. Additionally, the results assume access to an offline approximation algorithm for the specific stochastic problem, which may not exist or be efficient for all monotone problems. The assumption of independent random variables, while standard, restricts applicability to problems with correlated uncertainties.
This work contributes to the foundational theory of online learning in combinatorial settings. It provides a general framework for analyzing semi-bandit feedback in stochastic optimization, which can be applied to a wide range of problems in operations research and economics. The technique of using Hellinger distance with adaptive chain rules may inspire further theoretical work in bandit learning and statistical estimation. However, the impact is primarily academic and theoretical rather than industrial. The paper establishes near-optimal regret bounds for semi-bandit learning of monotone stochastic problems by leveraging Hellinger distance and adaptive chain rules. This theoretical advance improves upon prior bounds by a $\sqrt{n}$ factor and provides a unified framework for analyzing a wide class of stochastic optimization tasks, though it lacks empirical validation.
Recent research-level physics-reasoning benchmarks for LLMs reveal a failure mode that lies not in knowledge or execution, but in modeling the world being reasoned about. Models often silently alter given quantities, fabricate laws, or inject unstated assumptions to force a problem into familiar patterns, and then reason inside this corrupted world. We define this premise-level divergence as modeling drift: the gap between a model's internal axiomatic world model (AWM) and the true physical world specified in the problem. Although RLVR and PRM have yielded substantial gains in mathematics and code, their reward shapes do not directly supervise world modeling in axiomatic domains such as physics. We propose Reinforcement Learning with Axiomatic World Modeling (RLAWM), which optimizes the policy by penalizing divergence between its inferred axiomatic world model and the true physics world. RLAWM probes physics reasoning at two hierarchical levels (modeling and solving), each scored by a physically grounded reward. The modeling reward targets abstract knowledge formulation, evaluating whether the model captures the true axiomatic world before attempting calculation. The solving reward targets instantiated knowledge reasoning, checking whether the model's concrete derivation remains consistent with the axiomatic world. This two-level structure separates what world the model believes it is reasoning in from how it reasons within that world. The framework operates exclusively during training, with no extra cost during inference. RLAWM improves performance by up to +12.2 points on PhysReason and +5.2 on PHYSICS over the baseline. It also generalizes zero-shot across difficulty, source, and domain, with a 2.45x gain and +10 points over the strongest baseline on olympiad-level physics and math.
Primary: University of Florida
All Institutions: University of Florida, Northwestern University
The paper introduces a novel RL framework, RLAWM, that mitigates "modeling drift" in physics reasoning by decomposing the task into axiomatic world modeling and solving, using a dual reward system of motif-based fidelity scoring and executable code verification. By leveraging the verification-derivation asymmetry of physics, the method provides dense, physically grounded training signals that expand the reasoning boundary and improve out-of-distribution generalization, representing a substantial step forward in domain-specific LLM reasoning.
The paper proposes RLAWM, a reinforcement learning framework that explicitly targets "modeling drift" in physics reasoning. The core innovation is the decomposition of the reasoning process into a [modeling] block (externalizing the Axiomatic World Model) and a [solving] block (pseudocode derivation). The methodology introduces two distinct reward signals: a Modeling Reward ($R_M$) based on Physics Modeling Fidelity (PMF) using motif extraction (QEL, VCL, AEB) and an LLM-based soft matcher, and a Solving Reward ($R_S$) that translates pseudocode to executable Python and runs programmatic unit tests against a library of physics law templates. This approach leverages the verification-derivation asymmetry in physics, where checking invariants is easier than deriving them. The use of a frozen translator (Qwen2.5-Coder) to bridge natural language/pseudocode and executable code is a practical and effective design choice.
The experiments are rigorous and comprehensive. The authors evaluate on in-distribution benchmarks (PhysReason, PHYSICS) and out-of-distribution benchmarks (OlympiadBench, PHYBench, AIME) across three model families (Qwen3-VL, Gemma-3). Key results include a +12.2 point improvement on PhysReason and a significant reduction in "diversity collapse" (SubPass@k curves continue to improve with $k$, unlike RLVR baselines). Ablations clearly demonstrate the necessity of both $R_M$ and $R_S$, showing they target orthogonal failure modes (premise errors vs. execution errors). The analysis of the training signal density (reducing zero-gradient groups from 29.9% to 9.3%) provides strong evidence for the effectiveness of the dense reward signal.
High. The paper provides detailed descriptions of the reward components, the motif extraction schema, and the checker library construction. A GitHub repository is provided with code and models. Hyperparameters and compute requirements (4x B200 GPUs, ~6 hours) are specified. The use of standard models (GPT-4o-mini for extraction, Qwen for translation) ensures accessibility.
The method relies on a frozen LLM for motif extraction and a separate translator, adding training complexity and cost. The checker library is static and may not cover all physics domains equally well (noted lower performance in Quantum/Stat Mech compared to E&M/Optics). The approach is currently specialized for physics; while the authors suggest extension to math, the specific "axiomatic world model" structure may need adaptation for other domains.
This paper offers a significant conceptual shift in how LLMs are trained for scientific reasoning, moving beyond outcome-based rewards to process-based world-model alignment. The "modeling drift" framework is likely to influence future work in scientific LLMs. The technique of using executable code as a reward signal for reasoning steps is a robust pattern that could be adapted to other formal domains like mathematics or logic. The paper introduces a novel RL framework, RLAWM, that mitigates "modeling drift" in physics reasoning by decomposing the task into axiomatic world modeling and solving, using a dual reward system of motif-based fidelity scoring and executable code verification. By leveraging the verification-derivation asymmetry of physics, the method provides dense, physically grounded training signals that expand the reasoning boundary and improve out-of-distribution generalization, representing a substantial step forward in domain-specific LLM reasoning.
Despite breakthroughs in AI-driven molecular generation, existing discovery pipelines operate largely in silico, with wet-lab feedback reserved for validation rather than iterative improvement. Whether large language model (LLM)-based reasoning agents can accelerate de novo discovery by iterating with lab data remains unproven. Here we use protein-binding aptamer design as a sandbox (1-2 week cycles, \$1-\$3 per candidate) to demonstrate a wet lab in a closed loop with AI systems. We tasked agentic pipelines with developing aptamer binders against $α$-thrombin and myoglobin across three design-test-improve cycles ($\sim$1,000 candidates per round; $\sim$4,400 constructs). Statistically significant, bead-corrected on-target hit rates among new candidates rose across rounds from 0.0% in Round 1 to 26.0% in Round 2 and 42.1% in Round 3 on $α$-thrombin (37.1% excluding literature motif descendants), and from 10.2% in Round 1 to 3.7% in Round 2 and 33.5% in Round 3 on myoglobin. Single-clone bead-binding titrations of eight full-length $α$-thrombin candidates confirmed concordance with pooled sequencing enrichment (Pearson $r = 0.90$). In these assays, literature-derived motifs engineered by the agents to reduce off-target reactivity in the pooled screen reached apparent affinities of $K_D = 24.6$--$133$ nM, whereas de novo scaffolds showed low-micromolar binding ($K_D > 1\ μ$M). We also report how our pipelines used myoglobin binding data to exploit the G-quadruplex binder space, diagnose cross-reactivity as on-target specificity dropped from 75.0% to 19.1%, and recover by developing strategies that restored specificity to 72.4%. We demonstrate that empirical lab data is more valuable than existing computational proxies, showing how iterative experimental feedback enables AI agents to diagnose assay artifacts and refine molecular designs across rounds.
Primary: Google LLC
All Institutions: Google LLC
The paper demonstrates that LLM-based agentic pipelines can effectively iterate with wet-lab data to improve molecular design, achieving significant hit rates and nanomolar binders in a closed-loop aptamer discovery campaign. By rigorously benchmarking against computational predictors and showing their inadequacy, the study establishes the necessity of empirical feedback loops for AI-driven discovery in data-scarce biological domains, offering a robust framework for future self-driving laboratory systems.
The paper proposes a closed-loop framework for agentic biomolecular discovery, specifically using single-stranded DNA aptamers as a "sandbox" to test LLM-based reasoning agents. The methodology involves a multi-agent pipeline (Analyst, Strategist, Designer, Scorer) that iterates through design-test-improve cycles. The agents utilize a suite of computational tools (ViennaRNA, Boltz-2, ESMFold2-NA, etc.) and statistical analysis (DESeq2, OLS regression) to interpret wet-lab sequencing data. A key methodological contribution is the rigorous statistical framework for correcting bead-background noise and distinguishing specific binding from non-specific matrix adhesion, which the agents learn to apply autonomously. The approach demonstrates that agents can diagnose failure modes (e.g., G-quadruplex cross-reactivity) and adjust design strategies accordingly, moving beyond simple generation to iterative refinement based on empirical feedback.
The experiments are rigorous and well-controlled. The authors test two targets: alpha-thrombin (a positive control with known aptamers) and myoglobin (a de novo target). They report statistically significant improvements in hit rates across three rounds (0% to 42.1% for thrombin, 10.2% to 33.5% for myoglobin). Single-clone titrations confirm that the top candidates identified by the pooled screen exhibit nanomolar affinity, validating the high-throughput assay. The paper also includes a critical negative result: benchmarking 12 existing computational predictors shows they perform near random chance, highlighting the necessity of the wet-lab loop. The inclusion of inter-operator reproducibility checks (r=0.84-0.96) strengthens the reliability of the data.
The paper provides high reproducibility. It details the specific tools, versions, and licenses used. The statistical pipeline (DESeq2 + OLS regression) is clearly defined. The agent prompts and reasoning traces are partially shared in Table 1. However, the full agent code and specific prompt engineering details are not fully released, and the wet-lab protocol, while described, requires specialized equipment (MiSeq, magnetic beads, specific protein conjugates). The data availability statement promises release of full tables upon publication, which is standard but not yet fulfilled.
The primary limitation is the lack of a direct head-to-head comparison against non-agentic baselines (e.g., standard SELEX or Bayesian optimization) under identical budget constraints. The authors acknowledge this, stating that prospective trials with control arms are needed. Additionally, the study is limited to aptamers; generalization to more complex modalities like proteins or antibodies is suggested but not demonstrated. The cost and time per round ($1-$3 per candidate, 1-2 weeks) are low, but still higher than purely in silico methods, limiting the scale of search possible.
This paper is significant for the AI for Science community as it provides a concrete, validated example of LLM agents closing the loop with physical experiments. It challenges the assumption that computational predictors alone can drive discovery in data-scarce domains. The "sandbox" concept of using aptamers to test agentic workflows before applying them to harder targets is a valuable strategic insight. It may influence the design of future self-driving laboratories and AI co-scientist frameworks. The paper demonstrates that LLM-based agentic pipelines can effectively iterate with wet-lab data to improve molecular design, achieving significant hit rates and nanomolar binders in a closed-loop aptamer discovery campaign. By rigorously benchmarking against computational predictors and showing their inadequacy, the study establishes the necessity of empirical feedback loops for AI-driven discovery in data-scarce biological domains, offering a robust framework for future self-driving laboratory systems.
Differentially private learning aims to learn from data while protecting individual training examples. For binary classification, finite Littlestone dimension characterizes approximate private PAC learnability (Alon et al., JACM 2022). Understanding how this dimension controls private sample complexity remains a central question. For classes of Littlestone dimension $d$, the first general upper bound was doubly exponential in $d$ (Bun, Livni, and Moran, FOCS 2020). Subsequent work improved this to $\widetilde O(d^6)$ (Ghazi et al., STOC 2021) and then to $\widetilde O(d^5)$ (Lyu, STOC 2026). We prove a nearly linear upper bound of $O(d\log d)$. Our main algorithmic contribution is a structural exponential mechanism for data-dependent candidate sets equipped with a rank satisfying suitable stability conditions. For natural Littlestone classes such as subspace indicators, this rank has a simple geometric interpretation: it is the codimension. This yields a private PAC learner for subspaces of dimension at most $d$ over any field, with tight linear dependence on $d$ and no dependence on the ambient dimension, improving the quadratic dependence of Kaplan et al. (STOC 2025). Our second contribution is combinatorial: drawing on model-theoretic ideas, we construct a suitable rank for general Littlestone classes and establish the counting bounds required by the mechanism.
Primary: Technion -- Israel Institute of Technology
All Institutions: Technion -- Israel Institute of Technology, Google Research
The paper proves a nearly linear upper bound of $O(d \log d)$ on the sample complexity of differentially private PAC learning for classes of Littlestone dimension $d$, resolving a central open problem in the field. By introducing a structural exponential mechanism that leverages a model-theoretic rank, the authors achieve a tight dependence on the Littlestone dimension, significantly improving upon previous polynomial bounds and providing a new algorithmic framework for private learning.
The paper introduces a "structural exponential mechanism" that generalizes the classical exponential mechanism to handle data-dependent candidate sets. This is a significant algorithmic innovation in private learning theory. The core novelty lies in defining a combinatorial rank for hypotheses in Littlestone classes, derived from model-theoretic concepts (specifically Shelah's $2$-rank), and proving that this rank satisfies stability conditions necessary for differential privacy. The authors demonstrate that for subspace indicators, this rank corresponds to codimension, yielding a tight linear sample complexity bound $O(d \log d)$, which is optimal up to logarithmic factors. The methodology is rigorous, combining combinatorial arguments with privacy analysis.
This is a theoretical paper in learning theory and differential privacy. There are no empirical experiments, datasets, or benchmarks. The "evaluation" consists of mathematical proofs establishing upper bounds on sample complexity and verifying the privacy guarantees of the proposed mechanism. The contribution is entirely analytical.
As a pure theory paper, reproducibility is defined by the correctness of the proofs. The paper provides detailed technical overviews and formal definitions. However, the full proofs are truncated in the provided text (ending in Section 5.1), so a complete verification of the combinatorial lemmas (specifically Lemmas 5.2-5.4 and the counting bounds in Section 5.3) is not possible from the snippet alone. Assuming the standard rigor of the authors (Moran, Yan) and the venue (likely STOC/FOCS level), the theoretical framework is sound.
The primary limitation is that the result is a sample complexity bound, not a computational complexity bound. The paper explicitly states that it does not assert computational efficiency. The algorithm described (selecting candidates based on version spaces and ranks) may be computationally intractable for large $d$ or complex classes. Additionally, the bound is for the realizable setting; while reductions to the agnostic setting exist, they incur additional overhead. The log-star question regarding VC vs Littlestone dimension remains open.
This work resolves a major open problem in private learning theory by establishing a nearly linear dependence on Littlestone dimension. It provides a new algorithmic primitive (structural exponential mechanism) that can be applied to other private optimization problems involving data-dependent candidate sets. It bridges learning theory and model theory, offering new combinatorial tools for understanding the complexity of private learning. The result is significant for the theoretical understanding of the trade-offs between privacy and sample efficiency. The paper proves a nearly linear upper bound of $O(d \log d)$ on the sample complexity of differentially private PAC learning for classes of Littlestone dimension $d$, resolving a central open problem in the field. By introducing a structural exponential mechanism that leverages a model-theoretic rank, the authors achieve a tight dependence on the Littlestone dimension, significantly improving upon previous polynomial bounds and providing a new algorithmic framework for private learning.
Understanding how stochastic gradient descent learns internal representations in wide neural networks trained on rich target functions is a central problem in machine learning. We study this question in two-layer neural networks learning target functions represented by networks of the same architecture, with arbitrary nonlinearity, and with network width and input dimension growing proportionally. In this regime, standard approaches for describing the learning dynamics require tracking a diverging number of overlaps, which measure correlations between learned and target features, leaving no obvious description in terms of a few order parameters. Here, we develop a mesoscopic description of learning by characterizing the distribution of overlaps. In the early stages of learning, we find that a Gaussian description of overlap statistics reduces the learning dynamics to a few collective variables that track learning of the target function's linear and quadratic components. We then estimate the timescale of specialization, when Gaussian closure breaks down and individual learned features begin to strongly correlate with specific target features. We find that specialization in extensive-rank networks is accelerated via a quadratic shielding mechanism, whereby learning the linear and quadratic components of the target function filters gradient fluctuations. We finally describe how the activation function determines the pattern of specialization and characterize the subsequent refinement of learned features. Together, these results show how collective learning gives way to individual feature discovery and clarify when learning in large neural networks admits a simple statistical description.
Primary: International School for Advanced Studies (SISSA)
All Institutions: International School for Advanced Studies (SISSA), The Abdus Salam International Centre for Theoretical Physics (ICTP)
The paper develops a mesoscopic theory for learning dynamics in extensive-width neural networks, deriving closed-form equations for the pre-specialization regime and predicting the timescale of feature specialization via a quadratic shielding mechanism. By bridging the gap between narrow-network order parameter analyses and infinite-width mean-field theories, it provides a rigorous framework for understanding how collective learning transitions to individual feature discovery in high-dimensional settings.
The paper introduces a "mesoscopic" description of learning dynamics for extensive-width (proportional width and input dimension) two-layer neural networks. It bridges the gap between the Saad-Solla approach (which fails in the extensive limit due to diverging order parameters) and mean-field/NTK theories (which often miss feature learning details or require infinite width). The core method involves tracking the distribution of overlaps between teacher and student neurons via their empirical moments. The authors derive a "Gaussian closure" regime where the dynamics close on a few collective variables (linear and quadratic components), and then analyze the breakdown of this closure to predict the onset of "specialization" (individual feature alignment). The derivation of the "quadratic shielding" mechanism, which explains how learning low-order components filters gradient noise to accelerate specialization, is a particularly insightful theoretical contribution.
The paper provides strong numerical validation of the theoretical predictions. It compares the derived ODEs for the order parameters ($Q_1, Q_2$) against simulations of SGD for various activation functions (ReLU, tanh, ELU, etc.). The agreement between the theoretical loss plateaus and the specialization timescales is precise. The experiments confirm the existence of the distinct learning phases (collective alignment, specialization, refinement) and validate the scaling laws for specialization time with respect to network width and activation function properties.
The paper is highly reproducible from a theoretical standpoint, as it provides detailed derivations in the appendices. The experimental setup is standard (teacher-student setup with Gaussian inputs), and the specific activation functions and network sizes are clearly defined. While code is not explicitly linked in the text provided, the mathematical framework is sufficient for replication.
The analysis is restricted to two-layer networks with fixed second-layer weights (committee machines). Extending this to deeper networks or trainable readouts is left for future work. The "Gaussian closure" is an approximation that breaks down during specialization, and the paper provides estimates for the transition rather than a fully closed dynamic description of the specialization phase itself. The assumption of online SGD (one-pass) limits direct applicability to multi-epoch training regimes common in practice, though the authors argue relevance to LLM pre-training.
This work contributes to the fundamental understanding of how neural networks learn representations in high-dimensional regimes. It provides a rigorous framework for analyzing feature learning that goes beyond the NTK and mean-field limits. The insights into "quadratic shielding" and the role of activation functions in specialization patterns could inform the design of better initialization schemes or learning rate schedules for wide networks. It connects statistical physics methods with modern deep learning theory, offering tools to predict training dynamics and generalization errors in the extensive-width limit. The paper develops a mesoscopic theory for learning dynamics in extensive-width neural networks, deriving closed-form equations for the pre-specialization regime and predicting the timescale of feature specialization via a quadratic shielding mechanism. By bridging the gap between narrow-network order parameter analyses and infinite-width mean-field theories, it provides a rigorous framework for understanding how collective learning transitions to individual feature discovery in high-dimensional settings.
Diffusion large language models (dLLMs) have emerged as a promising alternative to autoregressive language models, enabling non-autoregressive text generation through parallel decoding. Existing dLLMs can be broadly categorized into two families: masked diffusion models (MDMs) and uniform diffusion models (UDMs) In this work, we theoretically show that MDMs are easier to optimize due to their mask-based transition schedule but have lower computational expressivity, whereas UDMs enable iterative state revision and greater expressivity but are substantially harder to train. This trade-off raises a natural question: can we obtain the expressivity of iterative revision without giving up the simple, easy-to-optimize transition schedule of MDMs? Motivated by this analysis, we propose Lookahead Editing (\ours{}), which augments masked diffusion decoding with a rewritable proposal state. At each decoding step, the model performs several lookahead updates to iteratively refine future predictions in this proposal state before committing tokens to the masked sequence. This design introduces a rewritable proposal that uses additional compute to improve expressivity while keeping the simple mask-based transition schedule of MDMs. We fine-tune Qwen2.5-7B and Qwen3-8B to obtain LED-\texttt{7B} and LED-\texttt{8B}, respectively. Across eight benchmarks, LED SFT yields average gains of $+11.4$ and $+7.7$ points over standard MDM SFT. Moreover, using SFT alone, LED-\texttt{8B} outperforms LLaDA2.1-16B Q-mode, which is trained with continued pre-training, supervised fine-tuning, and reinforcement learning.
Primary: University of Cambridge
All Institutions: University of Cambridge, NVIDIA
The paper introduces Lookahead Editing, a method that enhances masked diffusion language models by incorporating a rewritable proposal state to improve computational expressivity while maintaining optimization stability. It provides rigorous theoretical proofs regarding the circuit complexity of MDMs vs UDMs and demonstrates significant empirical gains in reasoning tasks, outperforming larger models trained with more complex pipelines.
The paper proposes "Lookahead Editing" (LED), a decoding strategy for Masked Diffusion Language Models (MDMs) that introduces a rewritable proposal state. The core innovation is bridging the gap between the optimization simplicity of MDMs (which have a fixed, easy-to-learn transition schedule) and the computational expressivity of Uniform Diffusion Models (UDMs) (which allow iterative revision but are harder to train). Theoretically, the authors prove that MDMs are contained in $TC^0$ while UDMs can reach $NC^1$, justifying the need for iterative revision. The method involves a three-branch training setup (clean, masked, proposal) where the proposal branch is generated by a pretrained MDM, allowing the model to learn to correct errors in the proposal. At inference, the model performs $K$ lookahead steps to refine the proposal before committing tokens. This is a clever architectural/algorithmic modification that preserves the training stability of MDMs while gaining the benefits of iterative refinement.
The experiments are conducted on Qwen2.5-7B and Qwen3-8B models. The results show significant improvements over standard MDM SFT baselines (+11.4 and +7.7 points average across 8 benchmarks). Notably, LED-8B outperforms LLaDA2.1-16B, which uses a more complex post-training pipeline (CPT + SFT + RL), using only SFT. The speed-accuracy trade-off analysis demonstrates that LED offers a better Pareto front than existing dLLMs. The benchmarks cover math, science, and code, providing a comprehensive evaluation of reasoning capabilities.
The paper provides detailed pseudocode for both training and inference. The hyperparameters (block size, batch size, learning rate) are specified. However, the specific datasets used for SFT are listed as open-source, but the exact mixture ratios and preprocessing steps are not fully detailed in the provided text. The reliance on a pretrained MDM for generating proposals adds a dependency that must be managed, but the overall procedure is clearly defined.
The method requires an additional forward pass (or multiple, depending on $K$) for the proposal generation, which increases inference latency compared to standard MDMs, although the paper argues this is offset by the ability to use fewer decoding steps or higher throughput at matched accuracy. The theoretical analysis, while rigorous, relies on specific assumptions about the circuit complexity of transformers that may not hold for all practical implementations. The performance gains are observed on specific benchmarks (math/code), and generalization to other domains (e.g., creative writing, long-form generation) is not explicitly tested.
This work contributes to the growing field of diffusion-based language models, offering a practical way to improve their reasoning capabilities without abandoning the stable training dynamics of masked diffusion. The theoretical insights into the optimization-expressivity trade-off are valuable for the community. The ability to outperform larger, more complexly trained models with a simpler SFT recipe suggests that architectural/decoding innovations can be as impactful as data scaling or RL. The paper introduces Lookahead Editing, a method that enhances masked diffusion language models by incorporating a rewritable proposal state to improve computational expressivity while maintaining optimization stability. It provides rigorous theoretical proofs regarding the circuit complexity of MDMs vs UDMs and demonstrates significant empirical gains in reasoning tasks, outperforming larger models trained with more complex pipelines.
Significant efforts are currently being directed toward the implementation of large-scale quantum information processors, but realizing useful quantum advantage also requires the development of efficient quantum algorithms. While for many practical problems the dominant resource costs are associated with a small set of computational primitives, optimizing them in the quantum setting presents unique challenges, including a reversibility constraint absent from ordinary classical computation. Here we introduce an evolutionary framework in which large language models (LLMs) propose circuit designs and an independent evaluator verifies and scores them. We apply this framework to tasks spanning numerical computation, chemistry, optimization, quantum simulation, and number theory. In seven of the eight tasks, the discovered circuits significantly reduce both Toffoli and qubit counts relative to the state of the art. In the eighth, the search discovers a new reversible architecture for prime-field modular inversion. By handling reversibility differently, the architecture makes divsteps -- a classical algorithm previously impractical for reversible integer arithmetic -- viable and substantially improves the Toffoli-qubit trade-off for modular inversion. These results open the door for evaluator-guided LLM co-design of new algorithmic structures for quantum computation.
Primary: Harvard University
All Institutions: Harvard University, Anthropic, QuEra Computing
The paper presents a novel evolutionary framework using LLMs to discover state-of-the-art quantum circuits, achieving significant reductions in resource costs across multiple tasks and uncovering new reversible algorithms. This work bridges the gap between AI-driven code generation and rigorous quantum algorithm design, offering a powerful tool for future quantum computing research.
The paper introduces a robust evolutionary framework where Large Language Models (LLMs) propose high-level Python circuit designs, which are then verified by an independent, programmatic evaluator. This approach effectively addresses the "reversibility constraint" of quantum computation by allowing the LLM to reason about algorithmic structure rather than just gate sequences. The methodology is sound, leveraging the fact that coherent circuits on computational basis states can be simulated efficiently (tracking bits and phases), enabling rapid feedback loops. The use of MAP-Elites to maintain diversity in the search space is a strong technical choice that prevents premature convergence to local optima.
The experimental results are highly impressive, demonstrating state-of-the-art improvements in Toffoli and qubit counts across eight diverse tasks, including Reed-Solomon decoding and modular inversion. The discovery of a new reversible architecture for prime-field modular inversion using divsteps is a significant algorithmic contribution, not just an optimization of existing circuits. The benchmarks are rigorous, comparing against established baselines and providing detailed resource accounting.
The paper provides sufficient detail on the circuit model, cost metrics, and search strategy to allow for reproduction of the framework. However, the specific LLM prompts and the exact "OpenEvolve" implementation details are not fully open-sourced in the text, which may limit immediate reproducibility for those without access to the specific tooling or LLM versions (Claude Mythos 5).
The framework is bounded by the task specification and the computational basis representation, excluding optimizations that rely on changing the encoding itself (e.g., coset representations). The reliance on a specific LLM (Claude Mythos 5) and the proprietary "OpenEvolve" framework may limit accessibility. Additionally, the TQ-cost metric is a simplification that may not capture all hardware-specific constraints like depth or routing.
This work has significant implications for the field of quantum computing, offering a scalable path to discovering efficient quantum algorithms. It demonstrates that LLMs can contribute to fundamental algorithmic design in domains with complex constraints, potentially accelerating the development of fault-tolerant quantum computers. The approach could be extended to other areas of AI for Science where programmatic evaluation is feasible. The paper presents a novel evolutionary framework using LLMs to discover state-of-the-art quantum circuits, achieving significant reductions in resource costs across multiple tasks and uncovering new reversible algorithms. This work bridges the gap between AI-driven code generation and rigorous quantum algorithm design, offering a powerful tool for future quantum computing research.
We study uniform approximation of the Hölder ball $\mathcal H_1^s([0,1]^D)$ by fully connected ReLU networks. Except in the one-dimensional Lipschitz case, existing constructions remain separated from the information-theoretic efficiency limit. This article closes the gap for every $s>0$ and $D\ge1$: networks with width $O(N)$, depth $O(L)$, and parameter magnitude $O(1)$ attain uniform approximation error of order $(N^2L^2\log N)^{-s/D}$. No relation between width and depth is required. The construction uses new encoding and decoding mechanisms that exploit higher-order finite-difference bounds for local Taylor coefficients of the target function. As a statistical consequence, empirical risk minimizers over the corresponding clipped network classes attain the minimax regression rate $n^{-2s/(2s+D)}$ without additional logarithmic factors. The paper also gives a detailed exposition of the key approximation techniques, covering both established and new mechanisms.
Primary: Munich Center for Machine Learning (MCML)
All Institutions: Munich Center for Machine Learning (MCML), Department of Statistics
The paper closes the gap between existing ReLU network approximation constructions and the information-theoretic efficiency limit for Hölder smooth functions, establishing optimal rates for width, depth, and parameter magnitude. This is a significant theoretical contribution that resolves a major open problem in neural network approximation theory, providing a complete characterization of the capacity of ReLU networks for smooth function approximation.
The paper presents a rigorous theoretical framework for the uniform approximation of Hölder smooth functions by ReLU neural networks. The core innovation is a "higher-order point-fitting" mechanism that exploits the regularity of local Taylor coefficients to compress their storage within the network weights. By using higher-order finite differences and a segmented discrete Taylor-Newton recursion, the authors achieve an approximation error of $(N^2L^2\log N)^{-s/D}$ with bounded parameter magnitudes ($O(1)$). This closes the gap between existing constructions and the information-theoretic efficiency limit for all smoothness orders $s>0$ and dimensions $D\ge1$. The methodology is highly sophisticated, combining established techniques like bit-extraction and median smoothing with novel encoding strategies that preserve smoothness during the flattening of multivariate coefficient grids via Hermite interpolation bridges.
As a pure theory paper, there are no empirical experiments, datasets, or benchmark comparisons. The "evaluation" consists of mathematical proofs establishing upper bounds on approximation error and matching lower bounds via metric entropy arguments. The statistical consequence (minimax regression rates) is derived theoretically.
The paper is highly reproducible in the mathematical sense. It provides self-contained proofs for all key components, including simplified versions of standard techniques. The construction is explicit, and the bounds are tight. However, without code, practical implementation of these specific network architectures for real-world data is not directly demonstrated.
The results are limited to uniform approximation on the unit cube $[0,1]^D$ for Hölder classes. The lower bound relies on metric entropy, which may not capture all aspects of practical learning dynamics. The paper does not address the optimization landscape or the trainability of these networks, which is a significant gap between theoretical optimality and practical deep learning.
This work provides a definitive answer to a long-standing question in neural network approximation theory. It establishes that ReLU networks are optimal approximators for smooth functions in terms of width, depth, and parameter magnitude. This has implications for understanding the capacity of neural networks and the theoretical limits of nonparametric regression. It may influence future theoretical work on network architectures and the design of approximation schemes for other function classes. The paper closes the gap between existing ReLU network approximation constructions and the information-theoretic efficiency limit for Hölder smooth functions, establishing optimal rates for width, depth, and parameter magnitude. This is a significant theoretical contribution that resolves a major open problem in neural network approximation theory, providing a complete characterization of the capacity of ReLU networks for smooth function approximation.
Recursive self-improvement, in which AI systems build stronger AI systems, is one promising route to superintelligence. We study a weaker form that can be measured today: frontier agents post-training a small raw base model with public data, as a test of their research potential. We build Athanor, an open agent-native system for autonomous post-training research (AutoPostTrain) on real GPUs. We give the agents full research autonomy, low-friction engineering tools, and efficient starter code. The agents carry out the research themselves: they read datasets and evaluation results, build their training data, write and modify training algorithms, tune hyperparameters, and analyze the results to decide what to try next. They work in a sandbox and cannot use external teacher models. Athanor's services supply frozen evaluation, managed jobs, decontaminated data, a persistent research record, and a held-out suite the agents are unaware of. In a 104-hour campaign, a team of four Claude Opus 5.5 agents, each on an eight-GPU node and given no research advice, improves Qwen3-4B-Base from 15.9 to 67.3 on a six-task development suite and from 9.2 to 48.2 on a held-out suite, against 66.5 and 47.9 for the official Qwen3-4B-Instruct-2507. To our knowledge, this is the first demonstration that a team of frontier agents can autonomously post-train a raw base model into a cross-domain generalist model that matches or surpasses the official instruct version on average. The campaign's full record shows that the agents reach this level by recombining known methods and each other's models: SFT on public reasoning data, RL with a rollout change one agent introduced, weight merges across agents, and rounds of self-distillation. With execution handled, today's agents already have the engineering and research ability to conquer this task. We hope Athanor provides a foundation for scaling autonomous AI research.
Primary: ByteDance
All Institutions: ByteDance, Duke University, The University of Chicago, Virginia Tech
The paper presents Athanor, an agent-native system that enables frontier AI agents to autonomously post-train a base model to match official instruct versions. This is a significant step toward recursive self-improvement, demonstrating that current agents possess the engineering and research ability to execute complex ML research tasks when provided with the right tools and autonomy, thereby validating the potential for AI-driven acceleration of AI development.
The paper proposes "Athanor," a system designed to facilitate autonomous post-training research by frontier AI agents. The core methodological contribution is the creation of an "agent-native" environment that provides agents with full research autonomy, low-friction engineering tools, and efficient starter code. The system architecture includes specific services to support this autonomy: frozen evaluation, managed job scheduling, decontaminated data access, a persistent research record, and a held-out suite that the agents are explicitly unaware of. The methodology emphasizes that agents must perform the entire research loop—reading data, building training sets, writing/modifying algorithms, tuning hyperparameters, and analyzing results—without external teacher models or human research advice. This represents a shift from using agents as code generators to using them as autonomous researchers within a constrained, verifiable sandbox.
The experimental setup involves a 104-hour campaign where a team of four Claude Opus 5.5 agents, each operating on an eight-GPU node, attempts to post-train the Qwen3-4B-Base model. The results are striking: the agents improved the model's performance from 15.9 to 67.3 on a six-task development suite and from 9.2 to 48.2 on a held-out suite. These scores surpass or match the official Qwen3-4B-Instruct-2507 model (66.5 and 47.9, respectively). The paper claims this is the first demonstration of frontier agents autonomously post-training a raw base model into a cross-domain generalist that matches official instruct versions. The analysis of the campaign record indicates the agents succeeded by recombining known methods (SFT on public reasoning data, RL with specific rollout changes, weight merges, and self-distillation), suggesting that the agents' ability to orchestrate existing techniques is the key driver of success.
The paper describes the system components (services, sandbox, tools) in detail, which aids in understanding the framework. However, as a preprint with no citation count and no provided code repository or demo URL, immediate reproducibility is limited. The reliance on specific frontier models (Claude Opus 5.5) and proprietary infrastructure (ByteDance GPUs) may limit external replication. The "agent-native" system itself is described as open, but the specific prompts, starter code, and internal logic of the agents are not fully detailed in the abstract or provided text, which is a barrier to full reproduction.
The primary limitation is the reliance on a specific, high-cost frontier model (Claude Opus 5.5) for the agents, making the results dependent on the capabilities of that specific model version. The experiment is limited to a single base model (Qwen3-4B-Base) and a specific set of tasks, so generalizability to other model architectures or domains is not yet proven. The "104-hour" timeframe, while impressive, is still a finite and relatively short window for complex research, and the paper does not explore what happens if the agents are given significantly more time or resources. Additionally, the lack of a public code release at the time of review hinders community verification.
This paper has significant implications for the field of AI safety and the development of superintelligence. It provides a concrete, measurable step toward recursive self-improvement, demonstrating that current agents can execute complex research tasks that were previously thought to require human oversight. The "Athanor" system could serve as a foundational benchmark for evaluating the research capabilities of future AI models. It also highlights the potential for AI systems to accelerate the pace of AI research itself, which has profound ethical and strategic implications. The finding that agents can match official instruct models through autonomous recombination of known methods suggests that the bottleneck for AI progress may be shifting from algorithmic discovery to execution and orchestration. The paper presents Athanor, an agent-native system that enables frontier AI agents to autonomously post-train a base model to match official instruct versions. This is a significant step toward recursive self-improvement, demonstrating that current agents possess the engineering and research ability to execute complex ML research tasks when provided with the right tools and autonomy, thereby validating the potential for AI-driven acceleration of AI development.
Large language models (LLMs) reproduce a human working-memory signature on the N-back task: performance declines as memory load N increases, and errors concentrate on lures, items that repeat the current stimulus at a neighboring but incorrect serial position. We ask whether this shared behavioral signature reflects a shared mechanism. The literature offers two candidate accounts. On a precision account, softmax attention cannot concentrate its weight on the item N positions back, so adjacent items are retrieved instead. On an interference account, attention selects the correct item, but the content of competing items corrupts the decision. We dissociate these accounts in 15 pretrained LLMs from three families (70 million to 14 billion parameters) and in minimal transformers trained from scratch. In all three families, the attention heads that carry the task attend to the target on most missed matches. Linear probes decode near-perfectly whether a matching letter occurred at the target or at a lure position, even on trials where the model false-alarms to the lure, although its own decisions barely distinguish the two. Sharpening attention yields negligible gains, whereas suppressing the letter identity of non-target items improves sensitivity (d') in nearly all model-by-load conditions. In larger models the task is spread over more heads, and removing interference helps more. In minimal transformers, precision limits a single attention layer, but a second layer removes this constraint. The working-memory limit of LLMs therefore lies in reading items out under interference rather than in locating them, consistent with interference accounts of human working memory.
Primary: Yale University
All Institutions: Yale University
The paper provides a rigorous mechanistic explanation for LLM working memory failures, demonstrating that the bottleneck lies in reading out items under interference rather than in locating them via attention. By combining theoretical analysis of minimal transformers with extensive empirical interventions on 15 pretrained LLMs, the authors successfully dissociate precision and interference accounts, offering a new perspective on the internal computations of large language models.
The paper employs a rigorous mechanistic interpretability framework to dissect the working memory limitations of Large Language Models (LLMs) on the N-back task. The methodology is sophisticated, combining behavioral analysis (signal detection theory, $d'$), representational probing (linear probes on residual streams), and causal interventions (activation patching, attention sharpening, and subspace suppression of distractor content). A key strength is the theoretical grounding provided for minimal transformers, where the authors derive propositions showing that single-layer attention is fundamentally limited by softmax dispersion (precision account), while two-layer architectures can overcome this via direct comparison. This theoretical foundation is then contrasted with empirical findings in large pretrained models, where the bottleneck shifts to interference during readout. The use of activation patching to identify task-critical heads and the specific intervention of suppressing letter-identity directions in non-target items are methodologically sound and directly address the core hypothesis.
The experimental scope is impressive, covering 15 pretrained LLMs from three distinct families (Pythia, GPT-2, Qwen2.5) spanning a 200-fold range in parameter size (70M to 14B). Additionally, the authors train 720 minimal transformers from scratch to isolate architectural constraints. The results are consistent across models: attention heads correctly locate the target even on error trials, and linear probes can decode the correct answer from activations even when the model's decision is wrong. The causal intervention results are particularly compelling; suppressing distractor content significantly improves sensitivity ($d'$), whereas sharpening attention yields negligible gains. The inclusion of recurrent baselines (LSTM, LRU, Mamba) further validates that the observed decline is not an artifact of softmax attention alone but relates to content-dependent retrieval mechanisms.
The paper provides detailed descriptions of the N-back task setup, prompt formatting, and intervention procedures. The use of standard, publicly available models (Pythia, GPT-2, Qwen2.5) enhances reproducibility. The minimal transformer training details (hyperparameters, seeds, architecture variants) are well-documented. However, the specific code for the subspace suppression intervention and the exact implementation of the linear probes are not provided in the text, which may pose a minor barrier for exact replication without access to the authors' codebase.
The study is limited to the letter N-back task with synthetic stimuli, which may not fully capture the complexity of real-world working memory tasks. The mechanistic analysis relies on a specific line-by-line prompt format with correct earlier answers shown, which might not generalize to all prompting strategies. The head-level account may not transfer to models larger than 14B, as the identified heads become less critical in larger models. Furthermore, the linear probes only demonstrate that information is decodable, not necessarily that it is the sole driver of behavior, although the causal interventions help mitigate this concern.
This work has significant implications for the field of mechanistic interpretability and the understanding of LLM limitations. It challenges the prevailing "precision" narrative regarding attention failures and highlights "interference" as a key bottleneck. This insight could guide future model architectures and training objectives aimed at improving working memory and reasoning capabilities. The findings also contribute to the broader debate on the convergence of neural network mechanisms and their analogy to human cognitive processes. The paper provides a rigorous mechanistic explanation for LLM working memory failures, demonstrating that the bottleneck lies in reading out items under interference rather than in locating them via attention. By combining theoretical analysis of minimal transformers with extensive empirical interventions on 15 pretrained LLMs, the authors successfully dissociate precision and interference accounts, offering a new perspective on the internal computations of large language models.
This paper shows that JEPA-style latent world models (LWMs) can learn across diverse domains and data sources by adopting the language-conditioned interface of video generation models. We establish this approach as an effective and efficient basis for planning with imagination and downstream policy fine-tuning. To this end, we introduce LaLa-World, a Language-conditioned Latent World model that learns from simulation and robot trajectories together with broader image and video data. Two designs make this work. First, we describe both numerical actions and visual changes in language, enabling joint training of a single future predictor across heterogeneous datasets. Second, we build the predictor on a language model, allowing it to interpret language conditions while predicting future latent states. The resulting predictor achieves higher average MPC planning performance than existing LWMs trained separately for each domain or task. Adding image and video data further improves average planning performance and robustness to action phrasing. In a matched comparison with an RAEv2-based video generation model baseline, LaLa-World reaches comparable planning performance with $6\times$ less training cost and $13.9\times$ faster planning. Beyond planning, the pretrained LaLa-World serves as an initialization for LaLa-Policy, which jointly predicts actions and future visual representations. LaLa-Policy performs competitively with Cosmos Policy on LIBERO and RoboCasa, with up to $5.5\times$ faster inference.
Primary: NAVER AI Lab
All Institutions: NAVER AI Lab
The paper introduces a novel language-conditioned latent world model that unifies heterogeneous control and video data, achieving superior planning efficiency and performance compared to both existing latent models and generative baselines. By leveraging an LLM backbone to interpret language actions, LaLa-World enables scalable training on web-scale data and facilitates efficient downstream policy learning, representing a significant step towards general-purpose world models.
The paper proposes LaLa-World, a latent world model that uses natural language as a unified interface for actions and visual changes. The core architectural innovation is replacing the traditional numerical action conditioning mechanism (e.g., AdaLN) with a pretrained Large Language Model (LLM) backbone (Qwen3.5-0.8B) that directly processes text conditions and predicts future latent states. This allows the model to ingest heterogeneous data: simulation/robot trajectories (where numerical actions are verbalized via templates) and web-scale image/video data (where captions/instructions serve as conditions). The method is theoretically sound, leveraging the LLM's semantic understanding to bridge the gap between diverse action spaces. The addition of a diffusion-based action head for policy fine-tuning (LaLa-Policy) follows recent trends in unified video-action models but applies them to latent representations rather than pixels.
The experimental evaluation is extensive and well-structured. The authors conduct a matched comparison between their latent model and a generative baseline (RAEv2 with DiT), demonstrating significant efficiency gains (6x less training cost, 13.9x faster planning) while maintaining or improving planning performance. They evaluate across seven control datasets and two zero-shot transfer domains (RoboCasa). The ablation studies are particularly strong, investigating the impact of training data composition (control vs. video), language backbone choice (Qwen vs. Gemma vs. BERT), and the sensitivity of the model to language phrasing. The "Communicable World Modeling" experiment, where an LLM uses the world model as a tool for planning, is a novel and compelling demonstration of the utility of the language interface.
The paper provides high reproducibility. Code and weights are released on GitHub. The authors specify the exact models used (Qwen3.5-0.8B, DINOv2/v3 encoders), the training data sources (OpenVid, OmniEdit, ShareGPT, DROID, etc.), and the hyperparameters for training and planning. The templates for converting numerical actions to language are described, and the evaluation protocols (CEM settings, success criteria) are detailed.
The primary limitation is the reliance on fixed templates for verbalizing numerical actions, which may limit the model's ability to generalize to actions not covered by the template vocabulary, although the video data helps mitigate this. The planning horizon is relatively short (max 6 steps), which may be insufficient for long-horizon tasks. The policy evaluation is limited to simulation (LIBERO, RoboCasa) and lacks real-robot validation. The model size (0.8B) is modest compared to state-of-the-art video generation models, and the authors acknowledge that scaling effects are not fully established.
This paper has significant potential impact on the field of embodied AI and world models. By demonstrating that latent world models can be scaled using web-scale video data via a language interface, it offers a path to more generalizable and data-efficient world models. The efficiency gains over generative models make latent planning more practical for real-time applications. The concept of "communicable world modeling" opens new avenues for human-AI interaction and LLM-agent integration. The paper introduces a novel language-conditioned latent world model that unifies heterogeneous control and video data, achieving superior planning efficiency and performance compared to both existing latent models and generative baselines. By leveraging an LLM backbone to interpret language actions, LaLa-World enables scalable training on web-scale data and facilitates efficient downstream policy learning, representing a significant step towards general-purpose world models.
Current state-of-the-art segmentation models still produce anatomically implausible shapes, such as broken vessels and fragmented glands, on the structures that are hardest to segment. Learned shape representations capture what structures look like and can help correct these errors, but most cover a fixed set of anatomical classes. Extending this coverage across classes and imaging modalities is costly because it often requires new voxel-level annotations. We propose Diffusion for Anatomical Reshaping with Text (DART), which learns a shared shape representation from segmentation probabilities. It uses anatomical descriptions to correct these probabilities without receiving medical images. Across 13 public benchmarks, absolute Dice gains reach 31% for individual gallbladder cases in CT and 10% for individual prostate cases in MRI. Using existing segmentation predictions, we scale a single shape representation from 50 to 377 anatomical classes without new voxel-level annotations for correction training. The shared representation also corrects four lesion classes absent from training and transfers from CT to MRI without retraining. Corrected labels also improve segmentation models, raising one model's Dice from 57.1% to 67.5%. The same pipeline supports annotations for 45,000 CT scans. These results suggest that sharing shape information within and across classes can reduce the need for new voxel-level annotations when extending segmentation to more anatomical classes and imaging modalities. The model is available at https://huggingface.co/Koushik45/DART-Seg
Primary: ELLIS Institute Finland
All Institutions: ELLIS Institute Finland, Johns Hopkins University
DART introduces a text-conditioned 3D diffusion model that learns a shared anatomical shape representation from segmentation probabilities, enabling the correction of segmentation errors across 377 classes and imaging modalities without new voxel-level annotations. The paper demonstrates rigorous evaluation on 13 benchmarks, showing significant Dice improvements and successful transfer from CT to MRI, while providing a scalable framework for generating high-quality training labels for medical segmentation models.
The paper proposes DART, a text-conditioned 3D diffusion model that learns a shared shape representation from segmentation probability maps rather than medical images. The core innovation is the use of anatomical text descriptions (generated by LLMs and reviewed by humans) to condition the diffusion process, allowing a single model to correct segmentation errors across 377 anatomical classes. The method utilizes SDEdit-style partial noising to refine predicted probabilities, preserving information that binary thresholding would discard. The approach is clever in that it decouples shape knowledge from image appearance, enabling transfer across modalities (CT to MRI) and classes without retraining. The reliance on text embeddings for class conditioning is a significant architectural choice that facilitates the scaling of the vocabulary.
The experimental evaluation is extensive, covering 13 public benchmarks. The paper demonstrates strong performance gains, with absolute Dice improvements reaching 31% for specific hard cases (gallbladder in CT) and 10% for prostate in MRI. A key strength is the rigorous comparison against baselines, including traditional post-processing (morphological closing), statistical shape models (ASM), and learned methods (Post-DAE, SegRefiner). The paper provides ablation studies on the conditioning mechanism (text vs. index) and the noise level selection. The transfer experiments to MRI and unseen lesion classes are particularly compelling, showing that the shape representation generalizes well. The downstream task of training segmentation models on corrected labels shows consistent improvements across 15 model configurations.
The authors commit to releasing code, model weights, anatomical descriptions, and per-class noise levels. The use of public datasets and standard evaluation metrics enhances reproducibility. The detailed description of the training data composition and the specific noise schedule parameters provides a clear path for replication. The release of annotations for 45,000 CT scans is a valuable resource for the community.
The method operates within the bounding box of the input prediction, meaning it cannot recover anatomy that is completely missing from the initial crop. The shape representation may inherit systematic errors from the source segmentation models used to generate the training probabilities. The performance on strongly atypical anatomy (e.g., severe tumors) is limited, though the paper notes a flagging mechanism for such cases. The reliance on LLM-generated descriptions introduces a potential source of bias or error if the anatomical descriptions are inaccurate, although human review mitigates this.
This work has significant potential to reduce the annotation burden in medical imaging by enabling the reuse of shape knowledge across classes and modalities. The ability to generate high-quality training labels from existing predictions could accelerate the development of segmentation models for rare or difficult anatomical structures. The shared shape representation could serve as a prior for other medical imaging tasks, such as registration or anomaly detection. The release of a large-scale annotated dataset further amplifies the impact. DART introduces a text-conditioned 3D diffusion model that learns a shared anatomical shape representation from segmentation probabilities, enabling the correction of segmentation errors across 377 classes and imaging modalities without new voxel-level annotations. The paper demonstrates rigorous evaluation on 13 benchmarks, showing significant Dice improvements and successful transfer from CT to MRI, while providing a scalable framework for generating high-quality training labels for medical segmentation models.
World models enable embodied agents to anticipate action consequences, yet jointly modeling locomotion and manipulation across diverse embodiments remains challenging due to heterogeneous action spaces and limited multimodal supervision. We present Ropedia W0, the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. We curate Xp-Data, comprising 10k hours of egocentric human data and 3k hours of robot data with unified multimodal annotations and systematic quality control. A unified kinematic representation encodes locomotion through camera poses and manipulation through projected skeletons, with robot skeletons combining embodiment-specific forward kinematics with extensive manual keypoint annotation and correction. We also introduce Xp-Teleop, a long-horizon loco-manipulation benchmark based on high-quality dexterous humanoid teleoperation data. Evaluations demonstrate state-of-the-art performance across five cross-embodiment benchmarks (DreamDojo, $Ψ_0$, $ω$-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, from zero-shot inference to few-shot and full-data post-training. Scaling human pre-training data improves performance on both human and robot evaluations. After distillation, our model supports real-time, minute-long rollouts, enabling live VR interaction. Applications in policy evaluation, model-based planning, and world-action modeling further demonstrate the model's utility for embodied decision-making.
Primary: Unknown (Likely a major tech lab or consortium given the scale of data, e.g., NVIDIA, Google, or a major robotics startup, but affiliations are redacted/missing in the provided text)
All Institutions: Unknown
Ropedia W0 introduces a cross-embodiment world foundation model that unifies human and robot loco-manipulation through a shared kinematic representation, demonstrating that scaling human egocentric data significantly improves robot world model performance. The paper makes a substantial contribution by addressing the heterogeneity of action spaces in embodied AI, providing a rigorous evaluation across multiple benchmarks, and introducing a new long-horizon teleoperation benchmark, thereby establishing a new paradigm for leveraging human data in robotic world modeling.
The paper proposes Ropedia W0, a world foundation model for loco-manipulation. The core methodological contribution is the unification of heterogeneous action spaces (human locomotion vs. robot manipulation) into a single kinematic representation. Specifically, it encodes locomotion via camera poses and manipulation via projected skeletons. For robots, this involves combining embodiment-specific forward kinematics with manual keypoint annotation. The model is trained on Xp-Data, a massive dataset of 10k hours of human and 3k hours of robot egocentric video. The approach leverages the abundance of human data to pre-train the world model, hypothesizing that human loco-manipulation dynamics transfer to robots. This is a significant architectural and data-engineering choice, moving away from purely robot-centric world models to a cross-embodiment paradigm.
The experimental section is extensive, evaluating the model on five cross-embodiment benchmarks: DreamDojo, $\Psi_0$, $\omega$-HOME, Xp-Teleop (a new benchmark introduced in the paper), and AgiBot. The paper demonstrates state-of-the-art performance in video prediction quality. Crucially, it validates the scaling hypothesis: increasing human pre-training data improves performance on both human and robot evaluations. The introduction of Xp-Teleop, a long-horizon benchmark based on dexterous humanoid teleoperation, is a valuable contribution to the field. The model's ability to support real-time, minute-long rollouts after distillation is a strong practical result, enabling live VR interaction.
Reproducibility is likely limited due to the scale of the data (13k hours of video) and the complexity of the manual annotation pipeline for robot skeletons. While the project page is provided, the specific details of the "systematic quality control" and the exact architecture of the world model (e.g., transformer variants, latent space dimensions) would need to be in the full text/appendix to be fully reproducible by smaller labs. The reliance on proprietary or large-scale in-house data makes independent verification difficult.
The primary limitation is the heavy reliance on manual keypoint annotation and correction for robot skeletons, which is labor-intensive and may not scale easily to new embodiments without significant human effort. The model's performance on out-of-distribution embodiments or tasks not covered by the 13k hours of data is not fully explored. Additionally, the "unified kinematic representation" may struggle with highly dexterous or non-rigid manipulations that are difficult to capture with standard skeletons.
This work has high potential impact on the field of embodied AI and robotics. By demonstrating that human video can effectively pre-train robot world models, it offers a pathway to overcome the data scarcity problem in robotics. The introduction of a cross-embodiment benchmark (Xp-Teleop) and a scalable world model architecture could accelerate research in model-based planning and policy evaluation. The real-time capability opens doors for interactive simulation and VR-based robot training. Ropedia W0 introduces a cross-embodiment world foundation model that unifies human and robot loco-manipulation through a shared kinematic representation, demonstrating that scaling human egocentric data significantly improves robot world model performance. The paper makes a substantial contribution by addressing the heterogeneity of action spaces in embodied AI, providing a rigorous evaluation across multiple benchmarks, and introducing a new long-horizon teleoperation benchmark, thereby establishing a new paradigm for leveraging human data in robotic world modeling.
Transformers are widely assumed to exhibit token-decoding explainability. In particular, in vision transformer (ViT) classifiers, decoding individual patch tokens has become a common diagnostic tool due to its tendency to yield human-interpretable outputs. In light of the increasing popularity of this approach, an important question is whether token-decoding explainability is an inherent property of transformers that should be expected to hold across architectures, scales, and training objectives, or merely an artifact of some implicit bias. In this paper we systematically study token-decoding explainability across diverse ViT settings, spanning architectures, scales, and training objectives, using both controlled experiments and real-world models. We find that while some models indeed exhibit aligned and interpretable token decodings, others exhibit either inverted explainability maps or a complete lack of correlation with the image semantics. As we show, these tendencies toward different explainability behaviors are not intrinsic to the architecture and can vary arbitrarily with changes in the optimization landscape. We provide theoretical support for these observations by demonstrating a setting where a single fixed toy transformer architecture can provably implement the optimal classification rule in multiple ways, which differ in their patch-decoding explainability patterns. Our results show that token-decoding explainability is not an inherent property of trained transformers, but rather an outcome contingent on the problem setting and thus should not be a-priori assumed to hold. Code can be found in https://ronensc.github.io/Explainability-Illusion/ .
Primary: Technion - Israel Institute of Technology
All Institutions: Technion - Israel Institute of Technology, Google Research
The paper challenges the assumption that Vision Transformers inherently possess patch-decoding interpretability by demonstrating through empirical and theoretical means that this property is contingent on optimization rather than architecture. By showing that identical classification performance can be achieved with completely different or inverted interpretability maps, the work provides a critical cautionary tale for the field, suggesting that current interpretability diagnostics based on token decoding may be unreliable artifacts of specific training dynamics rather than intrinsic model properties.
The paper employs a rigorous dual-pronged methodology combining systematic empirical analysis with theoretical counterexamples. The empirical component involves training and evaluating a diverse suite of Vision Transformers (ViTs) across varying architectures, scales, and training objectives to test the robustness of patch-decoding interpretability. The theoretical component constructs a provable toy model demonstrating that a fixed transformer architecture can implement the same optimal classification function via multiple weight configurations that yield diametrically opposed interpretability maps (aligned vs. inverted). This combination of empirical breadth and theoretical necessity is methodologically sound and effectively addresses the core research question.
The experimental evaluation is comprehensive, covering both controlled synthetic settings and real-world models, including Vision-Language Models (VLMs). The paper successfully demonstrates that interpretability is not a stable property, showing cases where decoding maps are inverted or uncorrelated with semantics. The inclusion of VLMs adds significant relevance, as these are currently the dominant paradigm in vision AI. The results are surprising and challenge a widely held assumption in the community, which is a hallmark of high-impact empirical work.
The authors provide a public code repository, which is a strong positive for reproducibility. However, the full paper text provided is a skeleton of LaTeX file references rather than the rendered content. While the abstract and structure suggest a complete study, the lack of visible detailed hyperparameters or specific dataset splits in the provided text limits the immediate assessment of granular reproducibility details. The code link is the primary vehicle for verification.
The theoretical proof relies on a "toy transformer" architecture, which may not fully capture the complexity of large-scale, pre-trained industrial models. While the empirical section addresses this, the gap between the simplified theoretical model and real-world complexity remains a potential limitation in generalizing the theoretical insights. Additionally, the paper focuses on the *existence* of the illusion rather than providing a robust method to *detect* or *correct* it in arbitrary black-box models, leaving practitioners without a clear diagnostic tool.
This paper has high potential for broader impact by correcting a fundamental misconception in the AI interpretability community. If token-decoding is not inherently reliable, it invalidates many existing interpretability tools and analyses that rely on this assumption. This could lead to a re-evaluation of how interpretability is measured and reported in future research, potentially shifting focus toward more robust, architecture-agnostic interpretability metrics. It serves as a crucial "reality check" for the field. The paper challenges the assumption that Vision Transformers inherently possess patch-decoding interpretability by demonstrating through empirical and theoretical means that this property is contingent on optimization rather than architecture. By showing that identical classification performance can be achieved with completely different or inverted interpretability maps, the work provides a critical cautionary tale for the field, suggesting that current interpretability diagnostics based on token decoding may be unreliable artifacts of specific training dynamics rather than intrinsic model properties.
Synthesizing a monocular video from a new camera trajectory requires preserving the observed scene geometry and motion while completing regions that become newly visible. Feed-forward 4D reconstructors recover the structure of the observed scene, whereas video diffusion models provide a generative prior for completing unseen content. We introduce Holo4D, a video reshooting framework that uses reconstructed 3D trajectories as a holistic source-view 4D scaffold. We project these trajectories into the requested views and encode their geometry and source appearance to guide a pretrained video diffusion model. Together with the source video and target cameras, the scaffold grounds generation in reconstructed observations while leaving unsupported regions to the generative prior. Experiments demonstrate the value of trajectory-based conditioning for camera control and novel-view synthesis. The framework accommodates different 4D reconstructors without retraining the video generator, allowing it to benefit from future advances in source reconstruction. More details can be found in our project page.
Primary: University of Illinois Urbana-Champaign
All Institutions: University of Illinois Urbana-Champaign, Adobe Research
Holo4D introduces a modular framework that uses feed-forward 4D reconstruction trajectories as a geometric scaffold to guide video diffusion models for camera-controlled video synthesis. The paper makes a significant contribution by decoupling the 3D/4D reconstruction task from the generative video model, allowing for the integration of state-of-the-art 4D reconstructors without retraining the diffusion model. This approach effectively grounds the generative prior in observed scene geometry and motion, leading to improved consistency and controllability in novel view synthesis. The method's ability to leverage future advances in 4D reconstruction makes it a robust and forward-looking solution for video reshooting and generation tasks.
The paper proposes Holo4D, a framework that bridges feed-forward 4D reconstruction and video diffusion models. The core method involves using a 4D reconstructor to generate geometric trajectories of scene points from a source video. These trajectories are then projected into the target camera views to create a "holistic 4D scaffold." This scaffold encodes both geometry and source appearance, serving as a conditioning signal for a pretrained video diffusion model. The approach is modular, allowing the diffusion model to remain frozen while benefiting from improvements in the upstream 4D reconstruction module. This decoupling is a significant architectural choice that enhances the method's longevity and adaptability.
The experiments demonstrate the effectiveness of trajectory-based conditioning for camera control and novel view synthesis. The paper likely compares against baselines that use only 2D projections or static 3D structures, showing that the dynamic 4D scaffold provides superior grounding for motion and geometry. The evaluation likely includes metrics for geometric consistency and visual fidelity, as well as user studies or qualitative comparisons showing the model's ability to complete unseen regions while preserving observed structure.
The paper provides a clear description of the pipeline, including the use of standard 4D reconstructors and pretrained video diffusion models. The modular nature of the framework suggests that components can be swapped, which aids in reproducibility. However, specific details on the training of the conditioning encoder (if any) and the exact integration points with the diffusion model's attention layers would be crucial for full reproduction.
The method relies on the quality of the feed-forward 4D reconstructor; errors in the initial reconstruction will propagate to the generated video. Additionally, the computational cost of running a 4D reconstructor followed by a video diffusion model may be high. The method is also limited to scenes where 4D reconstruction is feasible, potentially excluding highly dynamic or non-rigid scenes that are difficult to reconstruct.
This work has significant implications for the field of video generation and 3D/4D reconstruction. By providing a robust way to ground video diffusion in 3D/4D geometry, it paves the way for more controllable and physically consistent video synthesis. The modular design encourages the community to develop better 4D reconstructors, knowing they can be directly plugged into existing video generation pipelines. This could accelerate progress in both 3D reconstruction and generative video models. Holo4D introduces a modular framework that uses feed-forward 4D reconstruction trajectories as a geometric scaffold to guide video diffusion models for camera-controlled video synthesis. The paper makes a significant contribution by decoupling the 3D/4D reconstruction task from the generative video model, allowing for the integration of state-of-the-art 4D reconstructors without retraining the diffusion model. This approach effectively grounds the generative prior in observed scene geometry and motion, leading to improved consistency and controllability in novel view synthesis. The method's ability to leverage future advances in 4D reconstruction makes it a robust and forward-looking solution for video reshooting and generation tasks.
Video diffusion models (VDMs) trained on internet-scale data encode a rich prior on appearance, motion, and physics that is difficult to replicate with explicit 3D rendering and costly 3D data capture. We introduce Video Diffusion Avatars (VDAs), a method that creates photo-realistic, multi-view consistent avatars from a single image by fine-tuning a VDM for precise control over camera viewpoint, facial expression, and upper-body pose. Departing from 3D morphable models (3DMMs), VDAs use a learned expression space that lifts the expressiveness ceiling of 3DMMs, e.g., for the tongue and mouth interior. However, learned expression codes exhibit heavy viewpoint bias, as our DISTRACT benchmark reveals. We show that this bias can be removed through disentanglement, enabling their use as driving signals for multi-view consistent avatars. Camera and upper-body pose are controlled jointly via proxy-mesh renderings, giving independent control over expression, pose, and viewpoint. A camera-triplet training strategy makes the approach efficient: our models train for 34 hours on a single A100 using only public multi-view data. VDAs outperform state-of-the-art single-image 3D avatars on novel-view reenactment, reducing face LPIPS by 29% and improving perceptual quality (JOD) by 66% on BecomingLit. A static 3DGS reconstruction probe further confirms that our generations are the most multi-view consistent. Built on a video prior, VDAs produce effects beyond the reach of explicit renderers, including physically plausible hair dynamics, shadows, reflections, and refractions.
Primary: University of Tokyo
All Institutions: University of Tokyo, Toyota Motor Corporation
The paper introduces Video Diffusion Avatars (VDAs), a method that fine-tunes video diffusion models to create photo-realistic, multi-view consistent 3D avatars from a single image with precise control over expression, pose, and viewpoint. By leveraging the rich priors of VDMs and addressing viewpoint bias through disentanglement, VDAs outperform existing methods in perceptual quality and consistency, enabling physically plausible effects like hair dynamics and lighting that are difficult to achieve with explicit 3D rendering.
The paper proposes Video Diffusion Avatars (VDAs), a method that leverages the rich priors of video diffusion models (VDMs) to generate controllable, multi-view consistent 3D avatars from a single image. The core innovation lies in replacing explicit 3D morphable models (3DMMs) with a learned expression space, which allows for higher expressiveness (e.g., tongue and mouth interior) but introduces viewpoint bias. The authors address this via a disentanglement strategy, enabling the learned codes to serve as robust driving signals. Control over camera and pose is achieved through proxy-mesh renderings, and a "camera-triplet" training strategy ensures efficiency (34 hours on a single A100). This approach effectively bridges the gap between the photorealism of diffusion models and the geometric consistency required for 3D avatars.
The experiments demonstrate significant improvements over state-of-the-art single-image 3D avatar methods. On the BecomingLit dataset, VDAs reduce face LPIPS by 29% and improve perceptual quality (JOD) by 66%. A static 3D Gaussian Splatting (3DGS) reconstruction probe confirms superior multi-view consistency compared to baselines. The paper also highlights qualitative advantages such as physically plausible hair dynamics and lighting effects (shadows, reflections) that are difficult to achieve with explicit renderers. The introduction of the DISTRACT benchmark to reveal viewpoint bias in learned expression codes is a valuable diagnostic contribution.
The paper provides specific training details (34 hours on A100, public multi-view data), which aids reproducibility. However, without explicit links to code or detailed hyperparameter tables in the provided text, full reproduction may be challenging. The reliance on fine-tuning a VDM suggests that access to a strong base model is critical.
The method requires fine-tuning a video diffusion model, which may be computationally intensive for non-experts despite the claimed efficiency. The reliance on proxy-mesh renderings for pose control might limit the range of extreme poses not well-represented in the proxy. The learned expression space, while more expressive than 3DMMs, may still struggle with very rare or out-of-distribution expressions.
This work has significant implications for the field of 3D human modeling, potentially shifting the paradigm from explicit geometric modeling to implicit diffusion-based generation. It enables new applications in virtual reality, film production, and social media where high-fidelity, controllable avatars are needed. The ability to generate physically plausible effects like hair dynamics and lighting from a single image is a major step forward in content creation. The paper introduces Video Diffusion Avatars (VDAs), a method that fine-tunes video diffusion models to create photo-realistic, multi-view consistent 3D avatars from a single image with precise control over expression, pose, and viewpoint. By leveraging the rich priors of VDMs and addressing viewpoint bias through disentanglement, VDAs outperform existing methods in perceptual quality and consistency, enabling physically plausible effects like hair dynamics and lighting that are difficult to achieve with explicit 3D rendering.
Large Language Model (LLM) agents can handle increasingly complex tasks, navigating long trajectories of observations and actions. As these trajectories grow, key-value (KV) cache I/O and storage become the dominant cost. Trajectories can also exceed the LLM's maximum context length, requiring control over active cache growth. Prior work uses context engineering to reduce cache size by trimming or summarizing history. However, modifying the trajectory prefix invalidates subsequent cached states and triggers costly re-prefilling of retained context. Task accuracy can also suffer when too little context is retained to limit re-prefill costs. We propose cache engineering as a new scheme. It allows agents to directly control their active KV cache without modifying the trajectory or re-prefilling retained context. Specifically, agents can proactively use drop to exclude selected cache from future attention and repos to adjust retained keys' position encodings. We formalize these cache operations and identify history dependence as a serving challenge: identical token sequences can yield different cache states under different cache operation sequences, making prefix matching insufficient for correct cache reuse. Thus, we co-design the AKV serving engine. It provides a stateless interface compatible with existing serving APIs, while internally managing stateful cache reuse, eviction, and restoration. Experiments show that a tool-response rolling-drop cache policy improves average benchmark scores by 1.7 points and reduces average cost by 27.7%, compared with conventional threshold-triggered summarization. With AKV, it achieves up to 1.2x the output throughput and a 1.5x speedup in P95 request latency. Recursive self-improvement of cache policies further reduces cost by 7.1% over expert design. Our code is at https://github.com/thu-nics/AKV.
Primary: Tsinghua University
All Institutions: Tsinghua University
The paper proposes cache engineering as a new scheme for LLM agents, enabling direct control over KV cache without trajectory modification, and introduces the AKV serving engine to handle stateful cache reuse. This represents a significant technical contribution to LLM inference optimization, offering substantial improvements in cost and latency for agent-based applications.
The paper introduces "Cache Engineering," a paradigm shift from context engineering (trimming/summarizing) to direct manipulation of the Key-Value (KV) cache. The core method involves two operations: `drop` (excluding specific cache entries from future attention) and `repos` (adjusting position encodings of retained keys). The authors formalize the challenge of "history dependence," where identical token sequences can result in different cache states depending on prior operations, breaking standard prefix-caching assumptions. To address this, they propose the AKV serving engine, which maintains a stateless external interface but manages stateful cache reuse, eviction, and restoration internally. The methodology is sound and addresses a critical bottleneck in long-horizon agent tasks.
Experiments demonstrate that a tool-response rolling-drop policy improves average benchmark scores by 1.7 points while reducing costs by 27.7% compared to summarization. The AKV engine achieves 1.2x output throughput and 1.5x P95 latency speedup. The inclusion of recursive self-improvement for cache policies, yielding an additional 7.1% cost reduction, is a strong empirical result. The evaluation covers both accuracy and efficiency metrics, providing a comprehensive view of the trade-offs.
The code is available at the provided GitHub repository. The paper provides sufficient detail on the cache operations and serving engine design to allow for reproduction. The use of standard benchmarks and clear metric definitions enhances reproducibility.
The approach relies on specific modifications to the attention mechanism and position encodings, which may not be directly compatible with all LLM architectures or inference frameworks without significant adaptation. The "history dependence" issue, while solved by AKV, adds complexity to the serving infrastructure. The paper focuses on text-based agents; applicability to multimodal agents with different cache structures is not explored.
This work has significant implications for the deployment of LLM agents in production environments where cost and latency are critical. By decoupling cache management from context modification, it enables more flexible and efficient agent designs. The AKV engine could become a standard component in LLM serving stacks, particularly for applications involving long trajectories. The paper proposes cache engineering as a new scheme for LLM agents, enabling direct control over KV cache without trajectory modification, and introduces the AKV serving engine to handle stateful cache reuse. This represents a significant technical contribution to LLM inference optimization, offering substantial improvements in cost and latency for agent-based applications.
Factual recall is a lookup: some representation inside the network serves as the key, and what that key is made of decides which queries can retrieve the fact, what an edit must cover, and what a retriever can read. We call it the factual access state. In autoregressive (AR) models the subject position enriches attributes largely independently of the relation, and relation-specific selection happens later, at the readout position. In masked diffusion language models (DLMs) we find the state is assembled differently, and much earlier: in the first attention layer the wording of the relation is routed into the position of whichever entity is visible and the early feed-forward layers look the fact up on this fused representation. The state therefore varies with the relation's surface form and with the query direction, so the queries about one fact occupy a region of state space rather than a point, and an operation on the fact succeeds only for queries that enter it. We treat the state as an interface to a DLM's factual knowledge, and put it to two uses. Writing to the state gives a knowledge editor: from one declarative statement, Multi-View Edit fits one shared value per direction across that region and installs it in the early-MLP block, preserving source-prompt reliability while substantially improving paraphrase and reverse question answering, where AR editors gain near zero. Reading it suggests a complementary use: the state at each entity position can serve as a relation-specific retrieval vector with no training, returning both pieces of evidence for two-entity queries on which pooled, prompted and lexical retrievers almost never return both.
Primary: The University of Tokyo
All Institutions: The University of Tokyo, LIGHTSPEED
The paper identifies the "factual access state" in masked diffusion language models, demonstrating that relation-specific factual lookup occurs in early layers, and leverages this mechanism to develop a multi-view knowledge editor and a training-free retrieval method that significantly outperform autoregressive baselines in paraphrase and reverse query generalization.
The paper proposes a mechanistic explanation for factual recall in Masked Diffusion Language Models (DLMs), identifying a "factual access state" formed in the first attention layer. Unlike Autoregressive (AR) models, where relation selection happens late, DLMs fuse relation wording into entity positions early. The authors validate this via mask-native causal tracing and attention surgery. They leverage this insight for two applications: (1) Multi-View Edit, a knowledge editing method that installs shared values across multiple query views (forward, reverse, paraphrased) in early MLPs, and (2) a training-free retrieval method using entity-position states as relation-specific vectors. The methodology is rigorous, moving from mechanism identification to practical application, with clear ablations isolating the contribution of multi-view fitting and directionality.
Experiments are conducted on LLaDA-8B and Dream-7B. The editing results show significant improvements in paraphrase and reverse question answering compared to AR editors (which gain near zero on reverse tasks) and single-view DLM editors. The retrieval experiments demonstrate that per-entity states outperform pooled and lexical baselines on two-entity queries, specifically in retrieving both pieces of evidence. The use of CounterFact, BAKE, and QAEdit provides a comprehensive evaluation of editing reliability, generalization, and locality. The causal tracing results are robust across models and mask directions.
The paper provides a GitHub link and detailed reproducibility statements, including hyperparameters, seed usage, and compute requirements. The code is available, and the methodology is clearly described. The use of standard benchmarks and open-source models (LLaDA, Dream) enhances reproducibility.
The editor assumes identifiable subject, relation, and object spans, which may not be available in arbitrary user statements. Free-generation transfer on QAEdit remains incomplete, with wrong answers still being the majority on paraphrased questions. The retrieval method is evaluated on a specific benchmark built from PopQA, and its generalizability to other domains is not fully established.
This work bridges interpretability and knowledge editing for DLMs, offering a new perspective on how these models access factual knowledge. The multi-view editing approach could be adopted for more robust knowledge updates in DLMs, and the retrieval method provides a lightweight, training-free alternative for evidence retrieval. The insights into the "factual access state" may inform future architectures and editing techniques for bidirectional models. The paper identifies the "factual access state" in masked diffusion language models, demonstrating that relation-specific factual lookup occurs in early layers, and leverages this mechanism to develop a multi-view knowledge editor and a training-free retrieval method that significantly outperform autoregressive baselines in paraphrase and reverse query generalization.
Long-context language models (LCLMs) power agents that act over long trajectories, from improving a codebase to running research experiments to forecasters that condition on lengthy histories. These applications draw on three distinct long-context capabilities: recall, where the answer is a fact present in the context; state tracking, where the outcome can be computed by performing a series of transformations to a specified state; and pattern learning, where the goal is to make a prediction by sampling from a distribution inferred from many observations. Existing benchmarks primarily focus on the first capability. In this work, we introduce LEAPS (Long-context Evaluation of Adaptive Pattern learning and State tracking), a collection of 12 novel datasets, spanning agentic workflows, behavioral modeling, and forecasting of market and world events. By construction, LEAPS is resistant to data contamination, extends to millions of input tokens, and uses reliable metrics. We evaluate 30 API and open models from 8K to 512K input tokens. LEAPS distinguishes models that perform near perfectly on recall tasks, finding that such models' state tracking and pattern learning abilities can vary significantly and degrade much more dramatically with increasing input lengths--even for pattern learning, where more observations should make learning a theoretically easier task. Moreover, thinking models often underperform their base counterparts, exhausting their generation budget or oversimplifying the task. We hope LEAPS can guide the development of the next generation of LCLMs.
Primary: Jane Street
All Institutions: Jane Street, Princeton University
The paper introduces LEAPS, a comprehensive benchmark for evaluating long-context language models on recall, state tracking, and pattern learning. It provides rigorous experimental evidence that current models struggle with state tracking and pattern learning at scale, offering critical insights for the development of next-generation LCLMs.
The paper introduces LEAPS, a benchmark suite comprising 12 datasets designed to evaluate three specific long-context capabilities: recall, state tracking, and pattern learning. The methodology is rigorous, emphasizing resistance to data contamination through procedural generation of synthetic data and the use of reliable metrics. The distinction between these three capabilities is a strong conceptual contribution, as existing benchmarks often conflate them or focus heavily on recall. The approach to extending evaluations to millions of tokens is methodologically sound, though the reliance on API models for the upper end of the scale introduces potential variability in evaluation conditions.
The experimental scope is extensive, evaluating 30 models (both API and open-source) across input lengths from 8K to 512K tokens. The results provide valuable insights, particularly the finding that models excelling in recall do not necessarily perform well in state tracking or pattern learning. The observation that "thinking" models often underperform their base counterparts due to budget exhaustion or oversimplification is a significant and counter-intuitive empirical finding. The degradation of performance with increasing context length, even in tasks where more data should theoretically help (pattern learning), is a critical result for the field.
The authors provide a strong reproducibility statement, releasing code for data generation and evaluation, as well as the original evaluation data in Parquet format (with exceptions for licensing). The use of synthetic data generation scripts ensures that the benchmark can be regenerated, further mitigating contamination risks. The exclusion of social media data due to licensing is a minor limitation but does not significantly impact the core reproducibility of the main findings.
The primary limitation is the reliance on API models for the longest context lengths, which may not be fully reproducible by other researchers due to cost and access constraints. Additionally, the benchmark focuses on specific types of long-context tasks (agentic, behavioral, forecasting), which may not cover all potential applications of LCLMs. The performance of open models at the 512K token limit may be limited by hardware constraints, potentially skewing the comparison if not all models can be tested at the maximum length.
This benchmark has high potential impact as it addresses a critical gap in the evaluation of long-context models. By distinguishing between recall, state tracking, and pattern learning, it provides a more nuanced view of model capabilities than existing benchmarks. The findings regarding the limitations of current models, particularly the degradation with length and the underperformance of thinking models, will likely guide future research directions in model architecture and training. The benchmark's resistance to contamination makes it a robust tool for ongoing evaluation as models evolve. The paper introduces LEAPS, a comprehensive benchmark for evaluating long-context language models on recall, state tracking, and pattern learning. It provides rigorous experimental evidence that current models struggle with state tracking and pattern learning at scale, offering critical insights for the development of next-generation LCLMs.
Tracking large human-motion datasets has become the dominant route to general, expressive humanoid whole-body control, but scaling across many motions is expensive. Current pipelines use model-free reinforcement learning, either merging per-motion specialists through lossy distillation, or brute-forcing a single policy with cluster-scale compute. SuperMimic amortizes learning across all motions with one differentiable world model of their shared dynamics, fit from logged transitions of a standard rigid-body simulator. The world model and the tracking policy train together in one loop, and backpropagation-through-time over short imagined rollouts gives the policy analytic gradients. Fast imagined rollouts and analytic gradients make policy learning efficient. In 3 hours on a single GPU, SuperMimic surpasses the return of a state-of-the-art motion tracker baseline trained for 7 days on 128 GPUs. It reduces GPU-hours by more than $1000\times$ and fails on only 25 of 129,514 benchmark motions, versus 179 failures for the baseline. After 9.8 hours on two GPUs, SuperMimic achieves zero failures. From 1 to 264 hours of motion, the world model's prediction error stays stable, and GPU-hours to a target return grow slowly. Our policy transfers to a physical Unitree G1 across diverse and acrobatic motions. SuperMimic removes compute as the bottleneck to scaling humanoid whole-body control, making it practical to measure motion-tracking scaling laws.
Primary: UC Berkeley
All Institutions: UC Berkeley, Amazon (concurrent appointments), BAIR Center for Humanoid Intelligence
SuperMimic introduces a differentiable world model approach that amortizes humanoid whole-body control learning across large motion datasets, achieving over 1000x compute efficiency compared to state-of-the-art model-free RL baselines while improving tracking success rates and enabling real-world robot transfer.
The paper proposes SuperMimic, a method that amortizes humanoid whole-body control learning across a large dataset of motions using a single differentiable world model. Unlike standard model-free RL which relies on massive parallel sampling, SuperMimic fits a world model from logged transitions of a rigid-body simulator and uses Backpropagation-Through-Time (BPTT) over short imagined rollouts to provide analytic gradients to the policy. The key innovation is the co-training of the world model and the tracking policy in a single loop, allowing the policy to leverage the shared dynamics of the motions rather than learning from scratch for each motion or relying on lossy distillation of specialist policies. This approach effectively turns the simulation environment into a differentiable function, enabling efficient gradient-based optimization.
The experimental results are exceptionally strong and clearly demonstrate the efficiency gains of the proposed method. The claim of surpassing a state-of-the-art baseline (trained for 7 days on 128 GPUs) in just 3 hours on a single GPU is a significant reduction in compute cost (over 1000x). The benchmark of 129,514 motions provides a robust testbed, and the reduction in failure rates (from 179 to 25, and eventually to 0 with slightly more compute) indicates superior generalization and stability. The inclusion of real-world transfer to a Unitree G1 robot validates the practical utility of the learned policies, moving beyond simulation-only results. The scaling laws analysis, showing stable prediction error and slow growth in GPU-hours as motion data increases, further supports the scalability of the approach.
The paper provides specific details on the hardware (single GPU, 128 GPU baseline) and timeframes, which aids in understanding the magnitude of the improvement. However, the full text provided is a skeleton with section placeholders, so specific hyperparameters, network architectures, and loss function details are not visible in this excerpt. The project website is provided, which likely contains code and further details. The use of standard rigid-body simulators and logged transitions suggests a reproducible pipeline, assuming the world model fitting procedure is well-documented in the full paper.
The method relies on a differentiable world model, which may struggle with non-differentiable aspects of physics or complex contact dynamics not captured by the rigid-body simulator. The performance is heavily dependent on the quality and coverage of the logged transitions. While the compute reduction is dramatic, the absolute time (3-10 hours) is still non-trivial for rapid iteration, though vastly improved over cluster-scale training. The generalization to motions outside the training distribution is not explicitly detailed in the abstract, though the large benchmark suggests robustness.
This work has significant implications for the field of robotics and reinforcement learning. By removing compute as the primary bottleneck for scaling humanoid control, it enables the exploration of larger motion datasets and more complex behaviors. The approach of using differentiable world models for amortized learning could be extended to other domains requiring control over diverse tasks. The ability to measure scaling laws for motion tracking opens new avenues for understanding the relationship between data size, model complexity, and control performance. SuperMimic introduces a differentiable world model approach that amortizes humanoid whole-body control learning across large motion datasets, achieving over 1000x compute efficiency compared to state-of-the-art model-free RL baselines while improving tracking success rates and enabling real-world robot transfer.
Despite recent advances in policy learning, long-horizon manipulation remains difficult because learned policies must avoid compounding errors while preserving future feasibility. While search-based planning can explicitly reason over such future consequences, it quickly becomes expensive in high-dimensional continuous action spaces. This raises a central question: how can long-horizon reasoning from search be transferred into policies in continuous action spaces? To this end, we propose DISCO: Distilled Tree Search over Continuous Actions. DISCO uses progressive-widening MCTS in physics simulation to generate long-horizon plans. During search, a learned policy proposes continuous actions, while a learned value function scores actions by their expected progress toward task completion. Using GPU-parallelized simulation, DISCO periodically distills MCTS-generated plans into the policy-value model, improving the action proposals and value estimates used in subsequent searches. Finally, the learned policy-value model can be directly deployed without the need for simulator-based search at test time. Across simulated and real-world rearrangement tasks, DISCO outperforms behavior cloning, reinforcement learning, and search-based baselines.
Primary: Carnegie Mellon University
All Institutions: Carnegie Mellon University, Stanford University
The paper introduces DISCO, a method that distills long-horizon reasoning from GPU-parallelized MCTS into a deployable policy-value model for continuous action spaces. It represents a strong contribution to robot manipulation by effectively bridging the gap between sample-efficient planning and scalable policy learning, offering a viable solution to the compounding error problem in long-horizon tasks.
The paper proposes DISCO, a framework that bridges the gap between sample-efficient search-based planning and scalable policy learning in continuous action spaces. The core innovation is the use of progressive-widening Monte Carlo Tree Search (MCTS) within a physics simulator, where a learned policy proposes actions and a learned value function scores them. Crucially, the system employs a distillation loop: MCTS-generated plans are periodically distilled back into the policy-value model. This iterative refinement allows the policy to internalize the long-horizon reasoning capabilities of the search process. The method relies on GPU-parallelized simulation to make the search tractable, which is a significant engineering contribution. The final deployed model is a standard policy-value network that operates without the simulator, offering a clear advantage in real-world deployment latency and robustness compared to online search methods.
The evaluation covers both simulated and real-world multi-object rearrangement tasks. The paper claims to outperform behavior cloning (BC), reinforcement learning (RL), and search-based baselines. The inclusion of real-world experiments is a strong positive, as many papers in this domain remain purely simulated. The comparison against search-based baselines is particularly relevant, as it isolates the benefit of the distillation process over raw search performance. However, the specific metrics for "long-horizon" success rates and the degree of error compounding reduction would need to be scrutinized in the full results section to confirm the magnitude of improvement.
The method relies on specific hyperparameters for the MCTS (progressive widening factors, exploration constants) and the distillation frequency. The use of GPU-parallelized simulation suggests a dependency on specific hardware or simulation frameworks (likely Isaac Gym or similar), which may limit immediate reproducibility for labs without such infrastructure. The project page is provided, which likely contains code, aiding reproducibility.
The primary limitation is the dependency on a high-fidelity physics simulator for the training phase. If the simulator-to-real-world gap is significant, the distilled policy may suffer from performance degradation, although the real-world results suggest this is mitigated. Additionally, the computational cost of GPU-parallelized MCTS during training can be substantial, potentially limiting the scalability to extremely high-dimensional action spaces or very long horizons beyond what was tested.
This work has significant implications for the field of robot learning, particularly for tasks requiring precise, long-horizon manipulation. By demonstrating that search-based reasoning can be effectively distilled into a reactive policy, it offers a pathway to combining the robustness of planning with the speed of learned policies. This could influence future approaches to hierarchical reinforcement learning and model-based control. The paper introduces DISCO, a method that distills long-horizon reasoning from GPU-parallelized MCTS into a deployable policy-value model for continuous action spaces. It represents a strong contribution to robot manipulation by effectively bridging the gap between sample-efficient planning and scalable policy learning, offering a viable solution to the compounding error problem in long-horizon tasks.
Humanoid robots operating in human-centric environments must coordinate the entire body to traverse tight, cluttered spaces while maintaining balance. Existing methods often struggle to jointly consider dynamic stability and collision avoidance from high-dimensional perceptual inputs. Moreover, for humanoids, safe behavior is inherently multimodal: the same obstacle may admit multiple valid traversals, which is difficult to represent with a single deterministic policy. To address these challenges, we propose SWAP (Steerable Whole-body Avoidance Policy), a hierarchical framework that separates collision-aware decision making from stabilization. A high-level, perception-conditioned diffusion policy generates multimodal, collision-aware motion proposals, while a low-level whole-body RL controller tracks these proposals and maintains dynamic balance. To efficiently train the high-level policy, we obtain scalable supervision via hindsight labeling: we roll out the low-level policy under randomized commands, place obstacles only in the space that each executed motion did not sweep, so the motion is collision-free by construction, and re-render the perception the robot would have seen. At deployment, Monte Carlo sampling with simple cost signals steers the diffusion model toward safe, task-aligned behaviors without heavy reward tuning. Experiments in simulation and on a real Unitree G1 humanoid show reliable avoidance with static and dynamic obstacles. Test-time steering further enables customization, improving targeted rewards (e.g., speed, orientation, or posture) by 34% on average while maintaining a high collision avoidance rate. Together, these results support a practical solution to humanoid whole-body collision avoidance with test-time steerability. Project website: https://iscoyizj.github.io/SWAP-steerable-humanoid-oa-website/
Primary: Carnegie Mellon University
All Institutions: Carnegie Mellon University, Amazon FAR
The paper presents a novel hierarchical framework for humanoid whole-body collision avoidance that effectively combines diffusion-based multimodal planning with RL-based stabilization, demonstrating strong real-world performance and test-time steerability. The use of hindsight labeling for efficient training and the separation of collision-aware decision-making from dynamic balance represent significant methodological contributions that address key challenges in safe humanoid navigation.
The paper proposes SWAP, a hierarchical framework for humanoid whole-body collision avoidance that decouples high-level decision-making from low-level stabilization. The high-level component utilizes a perception-conditioned diffusion policy to generate multimodal motion proposals, addressing the inherent multimodality of safe navigation (e.g., going left vs. right around an obstacle). A key methodological contribution is the use of hindsight labeling for training the high-level policy: by rolling out a low-level RL controller under randomized commands and placing obstacles only in the space *not* swept by the robot, the authors generate collision-free supervision signals without requiring a complex reward function or simulator-in-the-loop collision detection during training. This is a clever and efficient training strategy. At deployment, Monte Carlo sampling with simple cost signals steers the diffusion model, allowing for test-time customization of behavior (e.g., prioritizing speed or posture) without retraining. The separation of concerns is well-motivated, and the use of diffusion models for multimodal action generation in robotics is a timely and relevant architectural choice.
The experiments are conducted in simulation and on a real Unitree G1 humanoid robot, which is a strong and practical choice for demonstrating real-world applicability. The evaluation includes both static and dynamic obstacles, covering a range of challenging scenarios. The paper reports a 34% average improvement in targeted rewards (speed, orientation, posture) via test-time steering while maintaining a high collision avoidance rate. While the specific metrics for "collision avoidance rate" and the baseline comparisons are not detailed in the provided text, the combination of simulation and real-robot validation, along with the quantitative improvement in steerability, suggests a rigorous experimental setup. The ability to demonstrate results on a physical humanoid is a significant strength, as many robotics papers remain purely in simulation.
The paper provides a project website, which likely contains code, videos, and additional details. The use of a standard platform (Unitree G1) and common simulation tools (implied) enhances reproducibility. The hindsight labeling technique is clearly described, allowing other researchers to implement similar training strategies. However, without access to the full code and hyperparameters, full reproducibility is difficult to assess. The reliance on a specific robot model may limit immediate reproducibility for labs without access to Unitree hardware, but the simulation components should be replicable.
The primary limitation is the reliance on a specific humanoid robot (Unitree G1) for real-world validation, which may not generalize directly to other platforms without significant adaptation. The method involves a two-stage process (high-level diffusion policy + low-level RL controller), which increases computational complexity and implementation difficulty compared to end-to-end approaches. The test-time steering via Monte Carlo sampling may introduce latency, which could be a concern in high-speed dynamic environments. Additionally, the paper does not extensively discuss failure modes or the limits of the approach in extremely cluttered or unpredictable environments.
This work has significant potential impact on the field of humanoid robotics and safe robot navigation. The hierarchical approach combining diffusion policies for multimodal planning with RL for stabilization is a promising paradigm that could be adapted to other robotic systems. The hindsight labeling technique offers a scalable way to generate training data for collision avoidance, which is a common challenge in robotics. The test-time steerability feature is particularly valuable for human-centric environments where user preferences or task-specific requirements may vary. This work could influence future research on safe robot operation, multimodal policy learning, and the integration of generative models in robotics. The paper presents a novel hierarchical framework for humanoid whole-body collision avoidance that effectively combines diffusion-based multimodal planning with RL-based stabilization, demonstrating strong real-world performance and test-time steerability. The use of hindsight labeling for efficient training and the separation of collision-aware decision-making from dynamic balance represent significant methodological contributions that address key challenges in safe humanoid navigation.
Video models provide rich priors for generalizable robotic manipulation, but plausible visual futures are not always executable. We introduce DreamCan, a hierarchical policy that grounds video models through sparse visual planning: a high-level planner adapted from a pre-trained video model and a low-level visuomotor policy. Rather than predicting dense trajectories, the high-level planner generates sparse visual subgoals that balance reachability and task utility. From offline, action-free robot videos, we estimate these quantities and use dynamic programming to select subgoal sequences without manual annotations. The resulting subgoals are more reliably attained than uniformly spaced alternatives, and the full system outperforms language, dense-video, and uniform-subgoal planners, reaching high planning success rates on Language-Table tasks and 91.2% success on LIBERO-Long with a 0.1B-parameter low-level policy. Additional suboptimal videos and test-time compute further improve success. The system also performs challenging real-world bimanual tasks, including box folding and packing with unseen object combinations. More results are available at https://dreamcan.github.io/.
Primary: Stanford University
All Institutions: Stanford University, Harvard University, National University of Singapore, Northwestern University
DreamCan introduces a hierarchical policy that uses sparse visual subgoals from video models to enable long-horizon robot manipulation. The paper presents a rigorous evaluation showing that this approach outperforms dense and uniform planning baselines, achieving high success rates in simulation and real-world bimanual tasks, thereby demonstrating the viability of video-based sparse planning for complex robotic manipulation.
The paper proposes "DreamCan," a hierarchical policy that bridges the gap between generative video models and executable robot actions. The core innovation lies in the high-level planner, which is adapted from a pre-trained video model to generate *sparse* visual subgoals rather than dense trajectories. This is a significant architectural choice, as dense video predictions are often physically infeasible for current robot hardware. The method estimates subgoal reachability and task utility from offline, action-free videos, utilizing dynamic programming to select optimal subgoal sequences without manual annotation. This unsupervised discovery of subgoals is a strong methodological contribution, reducing the reliance on expensive human demonstrations for long-horizon tasks. The low-level policy is a goal-conditioned visuomotor policy that executes the actions to reach these subgoals.
The experimental evaluation is robust, covering both simulation (Language-Table, LIBERO-Long) and real-world bimanual tasks. Achieving 91.2% success on LIBERO-Long with a small 0.1B-parameter low-level policy is a notable result, suggesting that the sparse planning approach effectively offloads complexity from the low-level controller. The comparison against language-based, dense-video, and uniform-subgoal planners provides a comprehensive baseline analysis. The inclusion of real-world bimanual tasks (box folding, packing) with unseen object combinations demonstrates generalization capabilities beyond simulation. The observation that additional suboptimal videos and test-time compute improve success rates is a valuable empirical finding for practitioners.
The paper provides a project URL (dreamcan.github.io) which likely contains code and additional results. The methodology relies on pre-trained video models and standard dynamic programming, which are well-understood components. However, the specific adaptation of the video model for sparse subgoal generation and the estimation of reachability/utility from action-free videos may require careful implementation details that are not fully visible in the abstract. The use of offline data is a practical advantage for reproducibility in real-world settings where collecting new action data is costly.
The reliance on pre-trained video models means the system's performance is bounded by the quality and coverage of the video priors. If the video dataset does not contain examples of the specific task or object configurations, the planner may fail. The method is designed for manipulation tasks; its applicability to locomotion or other robotic domains is not discussed. The computational cost of the high-level planner (video model inference) and the dynamic programming step could be a bottleneck for real-time operation, although test-time compute scaling is mentioned as a positive factor.
This work has significant implications for the field of robot learning, particularly for long-horizon manipulation tasks. By leveraging video models as priors for planning, it offers a path towards more generalizable and data-efficient robot policies. The sparse subgoal approach is a practical solution to the "plausibility vs. executability" problem in video-based planning. This could influence future research in hierarchical reinforcement learning and the integration of generative models into robotic control loops. DreamCan introduces a hierarchical policy that uses sparse visual subgoals from video models to enable long-horizon robot manipulation. The paper presents a rigorous evaluation showing that this approach outperforms dense and uniform planning baselines, achieving high success rates in simulation and real-world bimanual tasks, thereby demonstrating the viability of video-based sparse planning for complex robotic manipulation.
Differences in hand geometry, joint parameterization, and kinematic topology pose challenges for cross-embodiment dexterous grasp generation. However, spatial relations among links alone are insufficient to fully capture the hand's kinematic topology. We propose DexGraspFlow, a conditional link-level $\mathrm{SE}(3)$ flow-matching framework integrating geometric relations and kinematic topology. The framework encodes local object geometry, link geometry, and current poses. A dual-stream DiT models link--link coordination and link--object interaction, while attention biases derived from kinematic-tree distances provide explicit structural information for modeling evolving geometric interactions. The model jointly learns translational and rotational velocity fields, integrates them to generate link poses, and recovers joint configurations through inverse kinematics. Additionally, inference-time palm-orientation guidance enables controllable grasp generation without retraining. Across three robotic hands and 30 object--hand pairs, DexGraspFlow achieves a 95.33\% average simulation success rate, with an average generation latency, including inverse kinematics, approximately 49\% lower than that of a state-of-the-art baseline. Further experiments demonstrate zero-shot adaptability to seven finger-removal variants and palm-orientation control. A model trained on demonstrations for the deployed hand achieves a 91\% success rate over 100 real-robot grasping trials, supporting the framework's applicability to physical grasp execution.
Primary: Peking University
All Institutions: Peking University, University of Science and Technology of China
DexGraspFlow introduces a topology-aware SE(3) flow matching framework that significantly advances cross-embodiment dexterous grasp generation by explicitly encoding kinematic structure, achieving state-of-the-art success rates and efficiency with strong real-world validation.
The paper proposes DexGraspFlow, a framework for cross-embodiment dexterous grasp generation using SE(3) flow matching. The core methodological contribution is the integration of kinematic topology (via URDF-derived shortest-path distances) into the attention mechanism of a dual-stream DiT architecture. This addresses a specific gap in prior work (like TRO-Grasp) where spatial proximity was used as a proxy for kinematic relevance, which fails across different hand morphologies. The use of flow matching over diffusion models is a timely choice, offering faster inference (10 steps vs. 50+ for diffusion) and stable training. The formulation of separate translational (Euclidean) and rotational (SO(3) geodesic) velocity fields is mathematically sound and preserves the manifold structure of rotations. The inference-time palm-orientation guidance via tangent-space velocity corrections is a clever, training-free extension that enhances controllability.
The experimental evaluation is rigorous and comprehensive. The authors benchmark against strong, recent baselines (TRO-Grasp, DRO-Grasp, GenDexGrasp) across three distinct robotic hands (Barrett, Allegro, ShadowHand) and 30 object-hand pairs. The success rate of 95.33% is state-of-the-art, outperforming the previous best (TRO-Grasp) by a significant margin. The latency improvement (~49% lower) is substantial and practically important. The zero-shot generalization to finger-removal variants is a strong test of the topology-aware representation, showing the model understands the structural changes. The real-robot experiments (91% success rate) provide crucial validation beyond simulation, although the sample size (100 trials) is modest. The ablation studies clearly demonstrate the necessity of both geometric and topology biases.
The paper provides detailed implementation specifics, including dataset splits, hyperparameters (learning rate, batch size, network dimensions), and evaluation protocols (Isaac Gym settings, force magnitudes). The use of standard datasets (CMapDataset, ContactDB, YCB) and open-source tools (Pyroki, Isaac Gym) enhances reproducibility. However, the code and model weights are not explicitly linked in the provided text (though typically available for such papers), and the specific "filtered CMapDataset split" might require clarification if not publicly released.
The primary limitation is the reliance on a fixed set of training hands; generalization to completely novel hand morphologies (e.g., a 3-fingered hand vs. 5-fingered) is not tested, only finger removals. The method does not explicitly enforce joint limits or collision avoidance during the flow generation, relying on post-hoc IK which can fail or produce unnatural poses if the generated link poses are infeasible. The real-robot evaluation is limited to a single hand type (Linker Hand) and a small number of objects.
This work has significant implications for the field of dexterous manipulation. By providing a fast, accurate, and topology-aware method for cross-embodiment grasp generation, it facilitates the transfer of manipulation skills across different robotic platforms. The flow matching approach is likely to become a standard for generative robotics tasks due to its efficiency. The training-free guidance mechanism offers a practical way to add controllability to generative models without retraining, a valuable technique for other robotic applications. DexGraspFlow introduces a topology-aware SE(3) flow matching framework that significantly advances cross-embodiment dexterous grasp generation by explicitly encoding kinematic structure, achieving state-of-the-art success rates and efficiency with strong real-world validation.
Humanoid loco-manipulation is particularly challenging because it requires coordinating locomotion, balance, and object interaction, making it difficult to learn from limited demonstrations or from scratch through Reinforcement Learning (RL). We introduce DiPLO (self-distilled Diffusion Planning for Loco-manipulation), a framework that combines a self-distilled diffusion planner with a low-level whole-body tracking policy. Our planner is pretrained via supervised learning on motions derived from human-to-humanoid retargeting, while the tracker is trained via RL to execute them. DiPLO centers around two complementary post-training stages. First, we introduce Self-Distillation Diffusion Policy Optimisation (SD:DPO) to remove the planner's dependence on strong task conditioning, which we identify as a crucial ingredient for effective pretraining, as downstream fine-tuning does not recover comparable performance when it is omitted.. We then fine-tune this distilled model with Diffusion-Augmented Proximal Policy Optimization (DA:PPO), a maximum-entropy RL method for diffusion models, to improve task success and align generated plans with the frozen tracker's execution capabilities. Together, these stages enable motion generation through state-history and goal conditioning alone, enabling 99% task success in pick-carry-place tasks, compared to just 45% for the pretrained model. The resulting planner also generates motions for unseen object goals, with further evaluations on kicking and pushing/pulling tasks. Finally, we deploy our pipeline on the hardware, using our DA:PPO planner-tracker on a Unitree G1, achieving repeated successful executions without any real-world fine-tuning.
Primary: Technical University of Munich (TUM)
All Institutions: Technical University of Munich (TUM), McGill University, INRIA, Institute for Advanced Study (IAS)
DiPLO introduces a self-distilled diffusion planning framework that enables humanoid robots to perform complex loco-manipulation tasks from limited demonstrations by decoupling motion generation from execution and using a two-stage post-training process to remove dependence on kinematic references. The method demonstrates strong generalization to unseen goals and successful real-world deployment, highlighting the potential of hierarchical learning approaches for embodied AI.
The paper proposes DiPLO, a hierarchical framework for humanoid loco-manipulation that decouples high-level motion planning (diffusion model) from low-level execution (RL tracking policy). The core methodological contribution is a two-stage post-training process: (1) Self-Distillation Diffusion Policy Optimization (SD:DPO), which distills a teacher model conditioned on kinematic references into a student model conditioned only on state history and goals, effectively removing the need for explicit task priors at deployment; and (2) Diffusion-Augmented Proximal Policy Optimization (DA:PPO), which fine-tunes the distilled planner using maximum-entropy RL to align generated plans with the frozen tracker's capabilities. The use of SO(3) diffusion for rotational features is a technically sound choice that respects the manifold geometry of rotations. The separation of planning and tracking allows the planner to learn diverse motion structures from limited data while the tracker learns robust feedback control via RL, addressing the mode-covering vs. mode-seeking trade-off inherent in generative models vs. RL.
The experimental evaluation is rigorous and comprehensive. The authors provide extensive ablations on pretraining data augmentation (SBTO vs. symmetry), the importance of task conditioning during pretraining, and the specific benefits of SO(3) diffusion. They compare motion-level planning against action-level diffusion baselines (PDP, BeyondMimic), demonstrating the superiority of the hierarchical approach for survival and task success. The generalization evaluation, where the robot achieves unseen goals up to 5 meters away, is a strong indicator of the method's robustness. The real-world deployment on a Unitree G1 humanoid, achieving 100% success in pick-carry-place tasks without real-world fine-tuning, provides strong evidence of practical viability.
The paper provides detailed descriptions of the diffusion processes (VP diffusion for Euclidean features, IG_SO(3) for rotations), the DA-MDP formulation, and the training objectives. However, specific hyperparameters, network architectures (DiT details), and code availability are not explicitly provided in the text (no GitHub link found in the provided text). The reliance on specific datasets (OmniRetarget, OMOMO) and simulation environments (mjlab) may limit immediate reproducibility for those without access to these resources.
The method relies on external motion capture for object pose estimation in the real-world setup, which is a significant practical limitation for general deployment. The planning frequency is limited to ~12.5 Hz due to diffusion sampling, which may be too slow for highly dynamic tasks requiring faster replanning. The evaluation is primarily focused on pick-carry-place and similar tasks; performance on more complex, multi-object, or highly dynamic interactions is not fully explored. The dependence on a frozen tracker means that if the tracker fails, the planner cannot adapt its strategy in real-time to compensate for execution errors beyond the replanning horizon.
This work contributes to the growing field of humanoid robotics by providing a scalable method for learning complex loco-manipulation skills from limited demonstrations. The insight that strong task conditioning is useful for pretraining but can be distilled away for deployment is a valuable principle for robot learning. The combination of diffusion models for planning and RL for tracking offers a robust paradigm for handling the non-stationarity and balance constraints inherent in humanoid control. DiPLO introduces a self-distilled diffusion planning framework that enables humanoid robots to perform complex loco-manipulation tasks from limited demonstrations by decoupling motion generation from execution and using a two-stage post-training process to remove dependence on kinematic references. The method demonstrates strong generalization to unseen goals and successful real-world deployment, highlighting the potential of hierarchical learning approaches for embodied AI.
Every robot already carries high-resolution contact sensors: its own joints. Legged robots and humanoids exploit them to walk, run, and track whole-body motions without vision, but there the contacts are large and forceful. Is joint sensing also enough for precise, contact-rich assembly, where the robot must find a millimeter-scale opening, mesh gear teeth, or catch a thread through light contact? We study this question by training blind assembly policies that observe only a short history of joint positions, velocities, and their own actions, without visual, force, torque, or tactile sensing. The poses of both the fixture and the part are unknown to the policy, and the grasp is randomized substantially. The policy must therefore assemble the parts through contact, sensed exclusively through its own proprioception. Trained with reinforcement learning entirely in simulation over randomized parts, grasps, and contacts, such policies learn peg insertion, gear meshing, and nut threading. The peg insertion policy transfers zero-shot to the real world and inserts pegs of different shapes and sizes in 73% of trials. Controlled comparisons with matched visual policies in simulation show that blind policies generalize far better to unseen part shapes and sizes, though they can be slower. Training such policies needs thousands of different assemblies, each simulated with its clearance intact, so we also release an open-source generator that builds them procedurally from exact convex pieces for MuJoCo; it is more accurate and faster than convex decomposition or MuJoCo's signed-distance collision. Videos and interactive simulations are at https://younghyopark.me/blind-proprioceptive-assembly/.
Primary: Massachusetts Institute of Technology
All Institutions: Massachusetts Institute of Technology
The paper demonstrates that proprioceptive sensing alone is sufficient for precise, contact-rich robotic assembly tasks like peg insertion, gear meshing, and nut threading, achieving zero-shot sim-to-real transfer and superior generalization to unseen parts compared to visual policies. By introducing a rigorous simulation framework with exact convex collision geometry and a novel contact-supervised RL training method, the authors provide a robust, sensor-light alternative to vision-based manipulation, significantly advancing the state of the art in blind dexterity and sim-to-real transfer for precision tasks.
The paper proposes a novel approach to precise robotic assembly using only proprioceptive sensing (joint positions, velocities, and actions), explicitly excluding vision, force, torque, or tactile sensors. The core method involves training Reinforcement Learning (PPO) policies in simulation with extensive domain randomization of part geometry, grasp, and contact properties. A key technical contribution is the "contact supervision" auxiliary loss, which trains the policy's latent representation to predict contact states (free, touching, aligned) without providing privileged information to the actor, thereby accelerating learning from sparse rewards. Additionally, the authors introduce "AssemblyGen," an open-source procedural generator for MuJoCo that creates exact convex collision geometries for tight-clearance assemblies (pegs, gears, nuts), addressing a significant simulation bottleneck where standard convex decomposition or SDF methods fail to preserve sub-millimeter clearances accurately.
The experimental evaluation is rigorous and comprehensive. The authors demonstrate that blind policies achieve near-perfect success rates on in-distribution tasks and, crucially, generalize significantly better to unseen part shapes and sizes compared to matched visual policies. The zero-shot sim-to-real transfer is a strong result, with the peg insertion policy achieving 73% success on real hardware without fine-tuning, including on out-of-distribution pegs. The ablation studies effectively isolate the contribution of the contact supervision loss and the collision geometry generator. The comparison against scripted spiral search baselines highlights the efficiency of the learned reactive search strategy.
The paper provides high reproducibility standards. It releases an open-source generator for the simulation assets, which is a major barrier in this field. Detailed hyperparameters, architecture descriptions, and training ranges are provided in the appendix. The use of standard simulators (MuJoCo) and common algorithms (PPO) further aids reproducibility. The specific hardware setup (Flexiv Rizon 4S, Robotiq 2F-85) is clearly defined.
The primary limitation is the assumption that the robot starts in contact with the fixture; the policy does not handle the initial approach phase from a distance. The behavior is local and cannot recover from large deviations (e.g., if the part slips out of the grasp). The real-world evaluation is limited to peg insertion; gear meshing and nut threading were only validated in simulation. The reliance on precise simulation of friction and contact dynamics, despite domain randomization, remains a potential source of sim-to-real gap for other tasks.
This work challenges the prevailing assumption that precise manipulation requires vision or force sensing. By demonstrating that proprioception alone is sufficient for millimeter-scale assembly, it opens new avenues for robust, low-cost robotic systems that do not require expensive sensors or complex calibration. The released simulation tools will likely benefit the broader robotics community working on contact-rich tasks. The finding that blind policies generalize better to unseen geometries than visual policies is a significant insight for the design of robust manipulation policies. The paper demonstrates that proprioceptive sensing alone is sufficient for precise, contact-rich robotic assembly tasks like peg insertion, gear meshing, and nut threading, achieving zero-shot sim-to-real transfer and superior generalization to unseen parts compared to visual policies. By introducing a rigorous simulation framework with exact convex collision geometry and a novel contact-supervised RL training method, the authors provide a robust, sensor-light alternative to vision-based manipulation, significantly advancing the state of the art in blind dexterity and sim-to-real transfer for precision tasks.
We introduce REKI, a 2.4B vision-language-action model (VLA) pretrained exclusively on public data that sets a new Pareto frontier between performance and model size: on LIBERO-Plus, RoboTwin Clean2Rand and RoboCasa365, it is outperformed only by models about twice its size. REKI consists of a vision-language model (VLM) followed by a flow-matching action expert, and is pretrained on a curated mixture of public robot, egocentric human and vision-language data. Additionally, we endow it with a memory that gives the policy access to its recent past at negligible cost: a few register tokens in the action expert summarise the output of the VLM at each frame and persist in the expert's cache. This memory improves success in simulation and on real robots, and makes oscillations between poses about four times rarer. REKI uses a simple model design and training paradigm, and is optimised to be efficient at training and inference. We release the full code across data curation, training, inference and evaluation, deployment, and all pretrained and finetuned model weights: https://github.com/kyutai-labs/reki.
Primary: Kyutai
All Institutions: Kyutai
REKI introduces an efficient 2.4B VLA model with a novel register-token memory mechanism that achieves state-of-the-art performance on key robotics benchmarks while maintaining a small model size. The combination of flow-matching action experts, public-data pretraining, and a lightweight temporal memory strategy provides a robust and reproducible baseline that advances the field of vision-language-action modeling by balancing performance with computational efficiency.
The paper proposes REKI, a 2.4B parameter Vision-Language-Action (VLA) model. The architecture consists of a standard Vision-Language Model (VLM) backbone followed by a flow-matching action expert. The primary methodological contribution is the introduction of a lightweight memory mechanism using "register tokens" within the action expert. These tokens summarize the VLM's output at each frame and persist in the expert's cache, allowing the policy to access recent history with negligible computational overhead. The model is pretrained on a curated mixture of public robot, egocentric human, and vision-language data. The approach is notable for its efficiency and the specific design choice of using register tokens for temporal context rather than heavier recurrent or transformer-based memory blocks.
The evaluation is conducted on standard robotics benchmarks: LIBERO-Plus, RoboTwin Clean2Rand, and RoboCasa365. The paper claims REKI sets a new Pareto frontier, outperforming models of similar size and being outperformed only by models approximately twice its size. Specific empirical findings include a four-fold reduction in pose oscillations due to the memory mechanism. The inclusion of real-robot evaluation strengthens the claims, though the full text provided is largely a structural outline, so specific numerical comparisons must be inferred from the abstract's strong claims.
The paper explicitly states that full code for data curation, training, inference, evaluation, deployment, and all pretrained/finetuned weights are released. This high level of open-sourcing significantly enhances reproducibility and community utility.
The provided text is a preprint skeleton with section headers but lacks the detailed body text for rigorous scrutiny of specific ablation studies or failure cases. The reliance on "public data" may limit performance compared to models trained on proprietary, high-quality datasets, though the paper argues for efficiency. The "register token" memory is described as having "negligible cost," but the exact latency trade-offs in high-frequency control loops need verification in the full text.
By providing a strong, efficient baseline that is fully open-sourced, REKI lowers the barrier to entry for VLA research. The memory mechanism could be adopted by other architectures to improve temporal consistency in robotic control. The focus on public data pretraining is significant for democratizing access to capable robot policies. REKI introduces an efficient 2.4B VLA model with a novel register-token memory mechanism that achieves state-of-the-art performance on key robotics benchmarks while maintaining a small model size. The combination of flow-matching action experts, public-data pretraining, and a lightweight temporal memory strategy provides a robust and reproducible baseline that advances the field of vision-language-action modeling by balancing performance with computational efficiency.